2026-2-2

PaperBanana: Automating Academic

Illustration for AI Scientists

Dawei Zhu12*, Rui Meng2, Yale Song2, Xiyu Wei1, Sujian Li1, Tomas Pﬁster2and Jinsung Yoon2

1Peking University, 2Google Cloud AI Research

https://dwzhu-pku.github.io/PaperBanana/

Despite rapid advances in autonomous AI scientists powered by language models, generating publication-

ready illustrations remains a labor-intensive bottleneck in the research workﬂow. To lift this burden, we

introduce PaperBanana, an agentic framework for automated generation of publication-ready academic

illustrations. Powered by state-of-the-art VLMs and image generation models, PaperBanana orchestrates

specialized agents to retrieve references, plan content and style, render images, and iteratively reﬁne

via self-critique. To rigorously evaluate our framework, we introduce PaperBananaBench, comprising

292 test cases for methodology diagrams curated from NeurIPS 2025 publications, covering diverse

research domains and illustration styles. Comprehensive experiments demonstrate that PaperBanana

consistently outperforms leading baselines in faithfulness, conciseness, readability, and aesthetics.

We further show that our method eﬀectively extends to the generation of high-quality statistical plots.

Collectively, PaperBanana paves the way for the automated generation of publication-ready illustrations.

Google Condential & ProprietaryGoogle Condential & Proprietary 13

Methodology Diagrams Statistical Plots

Figure 1

Examples of methodology diagrams and statistical plots generated by PaperBanana,

which show the potential of automating the generation of academic illustrations.

1. Introduction

Autonomous scientiﬁc discovery is a long-standing pursuit of artiﬁcial general intelligence (Ghahra-

mani,2015;Langley,1987,2024;Schmidhuber,2010). With the rapid evolution of Large Language

Corresponding author(s): dwzhu@pku.edu.cn, lisujian@pku.edu.cn, jinsungyoon@google.com

* This work was done while Dawei was a student researcher at Google Cloud AI Research.

arXiv:2601.23265v1 [cs.CL] 30 Jan 2026

PaperBanana: Automating Academic Illustration for AI Scientists

Models (LLMs) (Anthropic,2025;Comanici et al.,2025;Liu et al.,2024;OpenAI,2025b;Yang et al.,

2025a), autonomous AI Scientists have demonstrated the potential to automate many facets of the

research lifecycle, such as literature review, idea generation, and experiment iteration (Gottweis et al.,

2025;Lu et al.,2024;Luo et al.,2025). Yet scientiﬁc discoveries achieve their full value only through

eﬀective communication. Despite their proﬁciency in textual analysis and code execution, current

autonomous AI scientists struggle to visually communicate discoveries, especially for generating

illustrations (diagrams and plots) that adhere to the rigorous standards of academic manuscripts.

Among these illustration tasks, generating methodology diagrams represents a signiﬁcant chal-

lenge, demanding both content ﬁdelity and visual aesthetics. Prior endeavors for diagram generation

have predominantly adopted the code-based paradigm, leveraging TikZ (Belouadi and Eger,2024;

Belouadi et al.,2025), Python-PPTX (Zheng et al.,2025), or SVG to programmatically synthesize

diagrams. While eﬀective for structured content, these methods can encounter expressiveness limita-

tions when attempting to produce the intricate visual elements – such as specialized icons and custom

shapes – that are increasingly common in modern AI publications. Conversely, although recent image

generation models (Deepmind,2025;OpenAI,2025a;Team et al.,2025;Wu et al.,2025a) have

demonstrated advanced instruction-following capabilities and high-quality visual outputs, consistently

generating academic illustrations that meet scholarly standards remains a diﬃcult task (Zuo et al.,

2025). Specialized expertise required for professional illustration tools often constrains researchers’

ability to freely express complex ideas, forcing them to invest substantial manual eﬀort into craft-

ing ﬁgures. This creates a signiﬁcant bottleneck in the eﬀective visual communication of scientiﬁc

discoveries.

In this paper, we introduce PaperBanana, an agentic framework designed to bridge this gap by

automating the production of high-quality academic illustrations. Given a methodology description

and diagram caption as input, PaperBanana orchestrates specialized agents powered by state-of-

the-art VLMs and image generation models (e.g. Gemini-3-Pro and Nano-Banana-Pro) to retrieve

reference examples, devise detailed plans for content and style, render images, and iteratively reﬁne

via self-critique. This reference-driven collaborative workﬂow allows the system to eﬀectively master

the logical composition and stylistic norms required for publication-ready illustrations. Beyond

methodology diagrams, our framework demonstrates signiﬁcant versatility by extending to statistical

plots, oﬀering a comprehensive solution for scientiﬁc visualization.

To rigorously evaluate our framework and address the absence of dedicated benchmarks for

automated academic illustration, we introduce PaperBananaBench, a comprehensive benchmark

for methodology diagram generation. The benchmark comprises 292 test cases and 292 reference

cases curated from NeurIPS 2025 publications, spanning diverse research topics and illustration

styles. To assess generation quality, we employ a VLM-as-a-Judge approach for reference-based

scoring against human illustrations across four dimensions: faithfulness, conciseness, readability, and

aesthetics, with reliability veriﬁed through correlation with human judgments.

Comprehensive experiments on our benchmark demonstrate the eﬀectiveness of PaperBanana.

Our method consistently outperforms leading baselines across all four evaluation dimensions—

faithfulness (+2.8%), conciseness (+37.2%), readability (+12.9%), and aesthetics (+6.6%)—as well

as the aggregated overall score (+17.0%) for diagram generation. We further show that our method

also seamlessly extends to statistical plots. Collectively, our method paves the way for automating

the generation of academic illustrations (Examples shown in Figure 1). As a demonstration of its

capability, ﬁgures marked with in this manuscript were entirely generated using PaperBanana.

Additionally, we discuss intriguing settings including using our framework to enhance existing human-

created illustrations and using image generation models for statistical plot generation. To sum up,

our contributions are:

PaperBanana: Automating Academic Illustration for AI Scientists

•

We propose PaperBanana, a fully automated agentic framework that orchestrates specialized

agents to generate publication-ready academic illustrations.

•

We construct PaperBananaBench to assess the quality of academic illustrations, particularly

methodology diagrams.

•

Comprehensive experiments show that our workﬂow signiﬁcantly outperforms leading baselines,

showing promise for automating the generation of academic illustrations.

2. Task Formulation

We formalize the task of automated academic illustration generation as learning a mapping from a

source context and a communicative intent to a visual representation. Let

𝑆

denote the source context

containing the essential information, and

𝐶

denote the communicative intent that speciﬁes the scope

and focus of the desired illustration. The goal is to generate an image

𝐼

that faithfully visualizes

𝑆

while fulﬁlling the communicative intent 𝐶, formulated as:

𝐼=𝑓(𝑆, 𝐶).(1)

To further guide the mapping function, the input can be optionally augmented by a set of

𝑁

reference

examples

E={𝐸𝑛}𝑁

𝑛=1

. Each example

𝐸𝑛

serves as a ground-truth demonstration, deﬁned as a

tuple

𝐸𝑛=(𝑆𝑛, 𝐶𝑛, 𝐼𝑛)

, where

𝐼𝑛

is the reference illustration corresponding to the context

𝑆𝑛

and

communicative intent 𝐶𝑛. Integrating this, the uniﬁed task formulation becomes:

𝐼=𝑓(𝑆, 𝐶, E),(2)

where Edefaults to ∅when no examples are used (i.e., zero-shot generation).

Among various types of academic illustrations, this paper primarily focuses on the automated

generation of methodology diagrams, which requires interpreting complex technical concepts and

logical ﬂows from textual descriptions into high-ﬁdelity, visually pleasing illustrations. In this setting,

the source context

𝑆

is the textual description of the method (e.g., methodology sections), and the

communicative intent

𝐶

is the ﬁgure caption specifying the scope and focus (e.g., “Overview of our

framework”).

3. Methodology

In this section, we present the architecture of PaperBanana, a reference-driven agentic framework for

automated academic illustration. As illustrated in Figure 2,PaperBanana orchestrates a collaborative

team of ﬁve specialized agents—Retriever, Planner, Stylist, Visualizer, and Critic—to transform raw

scientiﬁc content into publication-quality diagrams and plots. (See Appendix Gfor prompts)

Retriever Agent. Given the source context

𝑆

and the communicative intent

𝐶

, the Retriever Agent

identiﬁes

𝑁

most relevant examples

E={𝐸𝑛}𝑁

𝑛=1⊂ R

from the ﬁxed reference set

to guide the

downstream agents. As deﬁned in Section 2, each example

𝐸𝑖∈ R

is a triplet

(𝑆𝑖, 𝐶𝑖, 𝐼𝑖)

. To leverage

the reasoning capabilities of VLMs, we adopt a generative retrieval approach where the VLM performs

selection over candidate metadata:

E=VLMRet 𝑆, 𝐶, {(𝑆𝑖, 𝐶𝑖)}𝐸𝑖∈ R (3)

Speciﬁcally, the VLM is instructed to rank candidates by matching both research domain (e.g., Agent

& Reasoning) and diagram type (e.g., pipeline, architecture), with visual structure being prioritized

over topic similarity. By explicitly reasoned selection of reference illustrations

𝐼𝑖

whose corresponding

PaperBanana: Automating Academic Illustration for AI Scientists

Figure 2

[Generated by , textual description to reproduce this diagram is presented in Appendix E.]

Overview of our PaperBanana framework. Given the source context and communicative intent, we

ﬁrst apply a Linear Planning Phase to retrieve relevant reference examples and synthesize a stylistically

optimized description. We then use an Iterative Reﬁnement Loop (consisting of Visualizer and Critic

Agents) to transform the description into visual output and conduct multi-round reﬁnements to

produce the ﬁnal academic illustration.

contexts

(𝑆𝑖, 𝐶𝑖)

best match the current requirements, the Retriever provides a concrete foundation

for both structural logic and visual style.

Planner Agent. The Planner Agent serves as the cognitive core of the system. It takes the source

context

𝑆

, communicative intent

𝐶

, and retrieved examples

as inputs. By performing in-context

learning from the demonstrations in

, the Planner translates the unstructured or structured data in

𝑆into a comprehensive and detailed textual description 𝑃of the target illustration:

𝑃=VLMplan (𝑆, 𝐶, {(𝑆𝑖, 𝐶𝑖, 𝐼𝑖)}𝐸𝑖∈ E ).(4)

Stylist Agent. To ensure the output adheres to the aesthetic standards of modern academic

manuscripts, the Stylist Agent acts as a design consultant. A primary challenge lies in deﬁning

a comprehensive “academic style,” as manual deﬁnitions are often incomplete. To address this, the

Stylist traverses the entire reference collection

to automatically synthesize an Aesthetic Guideline

covering key dimensions such as color palette, shapes and containers, lines and arrows, layout

and composition, and typography and icons (see Appendix Ffor the summarized guideline and

implementation details). Armed with this guideline, the Stylist reﬁnes each initial description

𝑃

into

a stylistically optimized version 𝑃∗:

𝑃∗=VLMstyle(𝑃, G).(5)

This ensures that the ﬁnal illustration is not only accurate but also visually professional.

Visualizer Agent. After receiving the stylistically optimized description

𝑃∗

, the Visualizer Agent

collaborates with the Critic Agent to render academic illustrations and iteratively reﬁne their quality.

The Visualizer Agent leverages an image generation model to transform textual descriptions into

visual output. In each iteration 𝑡, given a description 𝑃𝑡, the Visualizer generates:

𝐼𝑡=Image-Gen(𝑃𝑡),(6)

where the initial description 𝑃0is set to 𝑃∗.

PaperBanana: Automating Academic Illustration for AI Scientists

Critic Agent. The Critic Agent forms a closed-loop reﬁnement mechanism with the Visualizer by

closely examining the generated image

𝐼𝑡

and providing reﬁned description

𝑃𝑡+1

to the Visualizer.

Upon receiving the generated image

𝐼𝑡

at iteration

𝑡

, the Critic inspects it against the original source

context

(𝑆, 𝐶)

to identify factual misalignments, visual glitches, or areas for improvement. It then

provides targeted feedback and produces a reﬁned description

𝑃𝑡+1

that addresses the identiﬁed

issues:

𝑃𝑡+1=VLMcritic(𝐼𝑡, 𝑆, 𝐶, 𝑃𝑡).(7)

This revised description is then fed back to the Visualizer for regeneration. The Visualizer-Critic loop

iterates for

𝑇=

3rounds, with the ﬁnal output being

𝐼=𝐼𝑇

. This iterative reﬁnement process ensures

that the ﬁnal illustration meets the high standards required for academic dissemination.

Extension to Statistical Plots. The framework extends to statistical plots by adjusting the Visualizer

and Critic agents. For numerical precision, the Visualizer converts the description

𝑃𝑡

into executable

Python Matplotlib code:

𝐼𝑡=VLMcode(𝑃𝑡)

. The Critic evaluates the rendered plot and generates a

reﬁned description

𝑃𝑡+1

addressing inaccuracies or imperfections:

𝑃𝑡+1=VLMcritic(𝐼𝑡, 𝑆, 𝐶, 𝑃𝑡)

. The

same

𝑇=

3round iterative reﬁnement process applies. While we prioritize this code-based approach

for accuracy, we also explore direct image generation in Section 6. See Appendix G.2 for adjusted

prompts.

4. Benchmark Construction

The lack of benchmarks hinders rigorous evaluation of automated diagram generation. We address

this with PaperBananaBench, a dedicated benchmark curated from NeurIPS 2025 methodology

diagrams, capturing the sophisticated aesthetics and diverse logical compositions of modern AI papers.

We detail the construction pipeline and evaluation protocol below; dataset statistics are in Figure 3.

4.1. Data Curation

Collection & Parsing. We begin by randomly sampling 2,000 papers from the 5,275 publications

at NeurIPS 2025 and retrieving their PDF ﬁles. Subsequently, we utilize the MinerU toolkit (Niu

et al.,2025) to parse these documents, extracting the text of the methodology sections, and all the

diagrams and their captions in the paper.

Figure 3

[Generated by ] Statistics of the test set

of PaperBananaBench (totaling 292 samples).

The average length of source context / ﬁgure cap-

tion is 3,020.1 / 70.4 words.

Filtering. We then apply a ﬁltering stage to

ensure data quality. First, we discard papers

without methodology diagrams, yielding 1,359

valid candidates. Second, we restrict the aspect

ratio (𝑤:ℎ) to [1.5,2.5]. Ratios below 1.5 are

excluded as methodology diagrams typically re-

quire wider landscape layouts for logical ﬂows,

while ratios exceeding 2.5 are unsupported by

current image generation models. Including

such outliers would introduce bias in side-by-

side evaluations by revealing the human origin

of candidates. This yields 610 valid candidates,

each a tuple

(𝑆, 𝐼, 𝐶)

, where

𝑆

is the methodol-

ogy description,

𝐼

is the methodology diagram,

and 𝐶is the caption.

Categorization. To facilitate future analysis of generating diﬀerent types of diagrams, we further

PaperBanana: Automating Academic Illustration for AI Scientists

categorize the diagrams into four classes, based on visual topology and content: Agent & Reasoning,

Vision & Perception,Generative & Learning, and Science & Applications (see Appendix Cfor deﬁnitions).

Gemini-3-Pro is used to perform the categorization, assigning samples with hybrid elements to their

predominant category.

Human Curation. Finally, we conduct a human curation phase to guarantee the integrity and quality

of the dataset. Annotators are tasked with verifying and correcting the extracted methodology

descriptions and captions, validating the correctness of diagram categorizations, and ﬁltering out

diagrams of insuﬃcient visual quality (e.g., overly simplistic, cluttered, or abstract designs). Following

this rigorous process, 584 valid samples remain. We randomly partition these into two equal subsets:

a test set (

𝑁=

292) for evaluation and a reference set (

𝑁=

292) to facilitate retrieval-augmented

in-context learning.

4.2. Evaluation Protocol

We utilize VLM-as-a-Judge to assess the quality of methodology diagrams and statistical plots. Given

the inherent subjectivity in evaluating visual design, we employ a referenced comparison approach

where the judge compares the model-generated diagram against the human-drawn diagram to

determine which better satisﬁes each evaluation criterion.

Evaluation Dimensions. Inspired by Quispel et al. (2018), we evaluate diagrams on two perspectives.

Detailed rubrics for each dimension are provided in Appendix H.

•

Content (Faithfulness & Conciseness): Faithfulness ensures alignment with the source context

(methodology description) and communicative intent (caption), while Conciseness requires focusing

on core information without visual clutter.

•

Presentation (Readability & Aesthetics): Readability demands intelligible layouts, legible text,

no excessive crossing lines, etc. Aesthetics evaluates adherence to the stylistic norms of academic

manuscripts.

Referenced Scoring. For each dimension, the VLM judge compares the model-generated diagram

against the human reference given the context and caption. It determines Model wins,Human wins,

or Tie based on relative quality, which are then mapped to scores of 100, 0, and 50, respectively. To

aggregate scores into an overall metric, we follow the design principle that information visualization

must primarily “show the truth” (Mackinlay,1986;Quispel et al.,2018;Tufte and Graves-Morris,

1983). We employ a hierarchical aggregation strategy, designating faithfulness and readability as

primary dimensions, and conciseness and aesthetics as secondary. If primary dimensions yield a

decisive winner (i.e., winning both, or winning one with a tie), this determines the overall winner. In

case of a tie (e.g., each wins one, or both tie), we apply the same rule to the secondary dimensions.

This hierarchical approach ensures that content ﬁdelity and clarity take precedence over aesthetics

and conciseness.

5. Experiments

5.1. Baseline Methods and Models

We compare PaperBanana against three baseline settings: (1) Vanilla, directly prompting the image

generation model to generate diagrams based on the input context (methodology description and

caption); (2) Few-shot, building upon the vanilla baseline by augmenting the prompt with 10 few-shot

examples, where each example consists of a triplet (methodology description, caption, diagram) to

enable in-context learning for the image generation model; (3) Paper2Any (Liu et al.,2025), an agentic

PaperBanana: Automating Academic Illustration for AI Scientists

Table 1 |Main results on PaperBananaBench. Best score in each column is in bold.

Method Faithfulness ↑Conciseness ↑Readability ↑Aesthetic ↑Overall ↑

Vanilla Settings

GPT-Image-1.5 4.5 37.5 30.0 37.0 11.5

Nano-Banana-Pro 43.0 43.5 38.5 65.5 43.2

Few-shot Nano-Banana-Pro 41.6 49.6 37.6 60.5 41.8

Agentic Frameworks

Paper2Any (w/ Nano-Banana-Pro) 6.5 44.0 20.5 40.0 8.5

PaperBanana (Ours)

w/ GPT-Image-1.5 16.0 65.0 33.0 56.0 19.0

w/ Nano-Banana-Pro 45.8 80.7 51.4 72.1 60.2

Human 50.0 50.0 50.0 50.0 50.0

framework that generates diagrams to present high-level ideas of the papers, which is the closest to

our setting. For VLM backbone, we default to Gemini-3-Pro, while for image generation model, we

experiment with Nano-Banana-Pro and GPT-Image-1.5. (See Appendix Cfor more implementation

details.)

5.2. Evaluation Settings.

Evaluating the quality of generated diagrams demands strong visual perception and understanding

capabilities, particularly for the Faithfulness dimension, which requires accurately identifying and

interpreting subtle modules and connections. Hence, we employ Gemini-3-Pro as our VLM-based

Judge. To validate its reliability, we randomly sampled 50 cases (25 from vanilla and 25 from our

method) and conducted a two-fold validation process:

Inter-Model Agreement (Consistency). First, we verify that our evaluation protocol is robust and

model-agnostic. We evaluated the agreement between our judge (Gemini-3-Pro) and other distinct

VLMs (Gemini-3-Flash and GPT-5). Kendall’s tau correlations with Gemini-3-Flash across the four

dimensions (Faithfulness, Conciseness, Readability, Aesthetic) and their aggregation are 0.51, 0.60,

0.45, 0.56, and 0.55, respectively; correlations with GPT-5 are 0.43, 0.47, 0.44, 0.42, and 0.45,

respectively. This conﬁrms the consistency of our protocol across diﬀerent judge models1.

Human Alignment (Validity). Second, we verify that our VLM judge is a valid proxy for human

evaluation. We tasked two human annotators to independently perform reference-based scoring

on the same 50 samples using the same rubrics, followed by a discussion to reach consensus on

conﬂicting cases. Kendall’s tau correlations between Gemini-3-Pro and human annotations are 0.43,

0.57, 0.45, 0.41, and 0.45, respectively. These strong correlations demonstrate that our VLM-based

judge aligns well with human perception. (See Appendix Bfor more details.)

5.3. Main Results

Table 1summarizes the performance of ours and baseline methods on PaperBananaBench.Pa-

perBanana consistently outperforms leading baselines across all metrics. We attribute the poor

performance of GPT-Image in both vanilla and agentic settings to its weaker instruction following and

text rendering capabilities compared to Nano-Banana-Pro, which fails to meet the strict requirements

According to existing literatures (Cohen,2013;Hollander et al.,2013), a Kendall’s tau correlation exceeding 0.4 is

generally considered to represent relatively strong agreement

PaperBanana: Automating Academic Illustration for AI Scientists

Table 2

Ablation study on PaperBananaBench. The shaded row indicates the default setting of

PaperBanana. We systematically ablate each agent component to assess its contribution. The



symbol denotes the Random Retriever which randomly selects 10 examples instead of performing

semantic retrieval.

Module

Faithfulness ↑Conciseness ↑Readability ↑Aesthetic ↑Overall ↑

Retriever Planner Stylist Visualizer Critic

①✓ ✓ ✓ ✓ 3 iters 45.8 80.7 51.4 72.1 60.2

②✓ ✓ ✓ ✓ 1 iter 38.3 75.2 50.6 68.9 51.8

③✓ ✓ ✓ ✓ - 30.7 79.2 47.0 72.1 45.6

④✓ ✓ -✓- 39.2 61.7 47.9 67.4 49.2

⑤✓-✓- 37.3 62.7 51.1 65.6 48.3

⑥-✓-✓- 41.9 58.6 43.1 62.9 44.2

of academic illustration. Similarly, while Paper2Any also supports generating paper ﬁgures, it priori-

tizes the presentation of high-level ideas rather than the faithful depiction of speciﬁc methodological

ﬂows necessary for methodology diagrams. This objective mismatch leads to its underperformance in

our evaluation setting.

In contrast, PaperBanana achieves comprehensive improvements over the Vanilla Nano-Banana-

Pro baseline: Faithfulness (+2.8%), Conciseness (+37.2%), Readability (+12.9%), and Aesthetics

(+6.6%), contributing to a +17.0% gain in the Overall score. Regarding performance across categories,

Agent & Reasoning achieves the highest overall score (69.9%), followed by Scientiﬁc & Application

(58.8%) and Generative & Learning (57.0%), while Vision & Perception scores the lowest (52.1%).

We also conducted a blind human evaluation on a subset of 50 cases to compare PaperBanana

against vanilla Nano-Banana-Pro (See Appendix Bfor details). The average win / tie / loss rate of

PaperBanana from 3 human judges is 72.7% / 20.7% / 6.6%, respectively. This further validates that

our agentic workﬂow shows promising improvements in automated methodology diagram generation.

(See Appendix Figure 7for case studies)

Despite the progress, we note that PaperBanana still underperforms the human reference in

terms of faithfulness. We have included some failure analysis in Appendix Figure 10 to provide

insights into the challenges.

5.4. Ablation Study

To understand the contribution of each agent component, we conduct an ablation study, with results

presented in Table 2.

Impact of the Retriever Agent. We compare the semantic retriever with random and no-retriever

baselines (rows

④

–

⑥

in Table 2). Without reference examples as guidance, the no-retriever setting

signiﬁcantly underperforms in Conciseness, Readability, and Aesthetics, as the Planner defaults

to verbose, exhaustive descriptions. Moreover, lacking exposure to academic diagram aesthetics,

this setting produces visually less reﬁned outputs. Interestingly, the random retriever achieves

performance comparable to the semantic approach, suggesting that providing general structural and

stylistic patterns is more critical than precise content matching.

Impact of the Stylist and Critic Agents. Comparing rows

③

and

④

shows that the Stylist boosts

Conciseness (+17.5%) and Aesthetics (+4.7%) but lowers Faithfulness (-8.5%), as visual polishing

sometimes omits technical details. However, the Critic Agent (row

①

vs.

③

) eﬀectively bridges this

PaperBanana: Automating Academic Illustration for AI Scientists

Figure 4

[Generated by ] Vanilla Gemini-3-Pro vs.

PaperBanana for statistical plots generation.

Figure 5

[Generated by ] Coding vs. Image

Generation for visualizing statistical plots.

gap, substantially recovering Faithfulness. Additional iterations further enhance all metrics, ensuring

a balance between aesthetics and technical accuracy.

5.5. PaperBanana for Statistical Plots Generation.

PaperBanana operates by ﬁrst synthesizing a detailed description of the target illustration, then

visualizing it into an image. Unlike methodology diagrams that prioritize aesthetics and logical

coherence, statistical plots demand rigorous numerical precision, making standard image generation

models unsuitable. To address this, we demonstrate that by adopting executable code for visualization,

PaperBanana seamlessly extends to statistical plot generation.

Testset Curation. Following the task formulation in Section 2, we assess PaperBanana’s capability

to generate statistical plots from tabular data and brief visual descriptions. Since raw data of statistical

plots is rarely available in academic manuscripts, we repurpose ChartMimic (Yang et al.,2025b), a

dataset originally constructed for chart-to-code generation. This dataset primarily includes statistical

plots from arXiv papers and Matplotlib galleries, paired with human-curated Python code. Leveraging

Gemini-3-Pro, we extract the underlying tabular data from the code and synthesize a brief description

for each plot. Following rigorous ﬁltering and sampling (see Appendix D), we curate 240 test cases

and 240 reference examples, stratiﬁed across seven plot categories—bar chart, line chart, tree &

pie chart, scatter plot, heatmap, radar chart, and miscellaneous—and two complexity levels (easy

and hard). For evaluation, we adhere to the protocol detailed in Section 4, with prompts speciﬁcally

tailored to statistical plots.

Figure 4compares PaperBanana with vanilla Gemini-3-Pro on our curated test set. Our method

consistently outperforms the baseline across all dimensions, achieving gains of +1.4%, +5.0%,

+3.1%, and +4.0% in Faithfulness, Conciseness, Readability, and Aesthetics, respectively, resulting

in a +4.1% overall improvement. Notably, PaperBanana slightly surpasses human performance in

Conciseness, Readability, and Aesthetics while remaining competitive in Faithfulness, showcasing its

eﬀectiveness for statistical plot.

6. Discussion

6.1. Enhancing Aesthetics of Human-Drawn Diagrams

Given the summarized aesthetic guidelines

, an intriguing question arises: can these guidelines serve

to elevate the aesthetic quality of existing human-drawn diagrams? To explore this, we implement a

streamlined pipeline where Gemini-3-Pro ﬁrst formulates up to 10 actionable suggestions based on

PaperBanana: Automating Academic Illustration for AI Scientists

Figure 6

Example of enhancing aesthetics of

human-drawn diagrams.

the original diagram and

, which are then

executed by Nano-Banana-Pro to reﬁne the im-

age. We evaluate the results using our reference-

based protocol, comparing the reﬁned output

against the original human-drawn diagram.

Across the 292 test cases, the reﬁned diagrams

achieved a win / tie / loss ratio of 56.2% / 6.8%

/ 37.0% in aesthetics against their original coun-

terparts, showing that the summarized aesthetic

guidelines can indeed serve to elevate the aes-

thetic quality of existing human-authored dia-

grams. An illustrative example is provided in

Figure 6. More examples are provided in Ap-

pendixFigure 8.

6.2. Coding vs Image Generation for Visualizing Statistical Plots

For statistical plots, code-based approaches have demonstrated remarkable eﬃcacy, as evidenced

by Figure 4and prior studies (Chen et al.,2025;Goswami et al.,2025;Yang et al.,2024). Given

the advanced ﬁdelity and visual appeal of recent image generation models, we compare code-based

(Gemini-3-Pro) and image-generation-based (Nano-Banana-Pro) approaches for the Visualizer agent

in PaperBanana, as shown in Figure 5. Results reveal distinct trade-oﬀs: image generation excels

in presentation (Readability and Aesthetics) but underperforms in content ﬁdelity (Faithfulness

and Conciseness). Manual inspection shows that while image models faithfully render sparse plots,

they struggle with dense or complex data, exhibiting numerical hallucinations or element repetition

(Appendix Figure 9). Thus, hybridly using image generation for sparse visualizations and code for

dense plots may oﬀer the best balance.

7. Related Work

7.1. Automated Academic Diagram Generation.

Automated academic diagram generation remains a long-standing challenge (Rodriguez et al.,2023).

Prior work primarily adopts code-based generation using TikZ (Belouadi and Eger,2024;Belouadi

et al.,2025;Hsu and Eger,2023;Zhang et al.,2025) or Python-PPT (Pang et al.,2025;Zheng

et al.,2025) for programmatic synthesis. While eﬀective for structured content, these approaches

face expressiveness limitations in generating the intricate visual designs prevalent in modern AI

publications.

Recent image generation models have achieved remarkable progress in synthesizing high-ﬁdelity,

visually sophisticated ﬁgures (Deepmind,2025;OpenAI,2025a;Tang et al.,2026;Team et al.,2025;

Zuo et al.,2025), oﬀering a promising alternative. Concurrent to our work, AutoFigure (Anonymous,

2026)

transforms scientiﬁc content into symbolic representations before rendering them as images

using GPT-Image. In comparison, our method achieves broader generalizability through adaptive

retrieval and academic-style transfer, with greater extensibility by supporting both methodology

diagrams and statistical plots in a uniﬁed pipeline.

For evaluation benchmarks, quality assessment of auto-generated diagrams remains less explored.

Most closely related to PaperBananaBench is SridBench (Chang et al.,2025), which evaluates

2Currently, AutoFigure is still an anonymous preprint. We will update the full citation once it is oﬃcially published.

PaperBanana: Automating Academic Illustration for AI Scientists

automated diagram generation from method sections and captions across computer science and

natural science domains. We will report results once it is publicly available.

7.2. Coding-Based Data Visualization

While the inherent complexity of academic diagram generation has deterred pioneering research,

visualizing statistical data has garnered extensive attention since the rise of language models. Early

endeavors (Dibia and Demiralp,2019) employed LSTM-based models to convert JSON data into

Vega-Lite visualizations, followed by few-shot and zero-shot coding approaches (Dibia,2023;Gal-

imzyanov et al.,2025;Li et al.,2024;Tian et al.,2024) leveraging large-scale backbones such as

ChatGPT (OpenAI,2022). More recently, agentic frameworks have demonstrated remarkable progress

in coding-based data visualization (Chen et al.,2025;Goswami et al.,2025;Seo et al.,2025;Yang

et al.,2024), leveraging fundamental mechanisms such as test-time scaling (Snell et al.,2024)

and self-reﬂection (Shinn et al.,2023). While this paper is more focused on automated generation

of academic diagrams and plots, these agentic frameworks can be seamlessly integrated into our

Visualizer Agent to enhance its capability in translating detailed descriptions of desired plots into

robust Python code. Complementary to generation, recent eﬀorts have also explored reversing plots

back into their original code (Wu et al.,2025b;Yang et al.,2025b), challenging both the perception

and coding capabilities of VLMs.

8. Conclusion

This paper introduces PaperBanana, an agentic framework designed to automate the generation

of publication-ready academic illustrations. By orchestrating specialized agents—Retriever, Plan-

ner, Stylist, Visualizer, and Critic—our approach transforms scientiﬁc content into high-ﬁdelity

methodology diagrams and statistical plots. To facilitate rigorous evaluation, we presented Pa-

perBananaBench, a comprehensive benchmark curated from top-tier AI conferences. Extensive

experiments demonstrate that PaperBanana signiﬁcantly outperforms existing baselines in faith-

fulness, conciseness, readability, and aesthetics, paving the way for AI scientists to autonomously

communicate their discoveries with professional-grade visualizations.

9. Limitations and Future Directions

As a pioneering work, although PaperBanana achieves promising results, it inevitably faces certain

limitations. This section will discuss these limitations in detail, and outline the corresponding future

directions we envision.

Towards Editable Academic Illustrations. The most prominent limitation of PaperBanana lies in

the raster nature of its output. Unlike vector graphics—which are preferred in academic contexts

for their inﬁnite scalability and precise detail preservation—raster images are inherently diﬃcult to

edit. While generating outputs at 4K resolution serves as a viable workaround to ensure high visual

ﬁdelity, it does not fundamentally resolve the challenge of post-generation modiﬁcation. To address

this, we envision three potential solutions catering to varying levels of editing needs. For minor visual

adjustments, leveraging state-of-the-art image editing models, such as Nano-Banana-Pro, serves as the

most direct approach. For more structural modiﬁcations, a reconstruction pipeline as exempliﬁed by

Paper2Any (Liu et al.,2025) and Edit Banana (BIT-DataLab,2025) can be adopted: employing OCR

for text extraction and SAM3 for pattern segmentation, followed by reassembling these elements on

presentation slides (e.g., via Python-PPTX). While currently facing challenges when handling complex

backgrounds and intricate visual elements, we anticipate that training specialized element extraction

PaperBanana: Automating Academic Illustration for AI Scientists

models will signiﬁcantly enhance the robustness of this reconstruction. Finally, a more advanced

direction involves developing a GUI Agent capable of autonomously operating professional vector

design software (Huang et al.,2026;Sun et al.,2025), such as Adobe Illustrator. This would enable

the direct generation of fully editable vector graphics, although it necessitates the agent to possess

exceptional perception, planning and interaction capabilities.

The Trade-oﬀ between Style Standardization and Diversity. The second limitation lies in the

trade-oﬀ between style standardization and diversity. While our uniﬁed style guide ensures rigid

compliance with academic standards, it inevitably reduces the stylistic diversity of the output. Future

work could explore more dynamic style adaptation mechanisms that allow for a broader range of

artistic expressions and personalized aesthetic choices while maintaining professional rigor.

The Challenge of Fine-Grained Faithfulness. While PaperBanana excels in aesthetics, a perfor-

mance gap in faithfulness compared to human experts remains. As shown in our failure analysis

(Figure 10 in the Appendix), the most prevalent errors involve ﬁne-grained connectivity, such as

misaligned start/end points or incorrect arrow directions. These subtleties often escape the detection

of current critic models, limiting the eﬃcacy of self-correction. We posit that closing this gap primarily

hinges on advancing the ﬁne-grained visual perception capabilities of the foundation VLMs.

Advancing Evaluation Paradigms. Following existing practices, our evaluation adopts a reference-

based VLM-as-a-Judge setup. Despite its eﬀectiveness, we acknowledge that this evaluation paradigm

still faces inherent challenges. First, regarding faithfulness, quantifying structural correctness remains

challenging, as detecting subtle errors in connectivity and notation requires high-precision scrutiny.

Future protocols could beneﬁt from incorporating ﬁne-grained, structure-based (Liang and You,2025)

or rubric-based (Huang et al.,2026;Li et al.,2025) metrics, which may oﬀer higher accuracy despite

their increased computational complexity. Second, for subjective dimensions such as aesthetics, we

observe that textual prompting is often insuﬃcient to fully align the VLM with human preferences.

We envision that training customized reward models to bridge this alignment gap represents a crucial

direction for future research.

Test-Time Scaling for Diverse Preferences. Currently, our framework produces a single output

for each query. However, given the inherent stochasticity of generative models and the subjectivity

of aesthetic preferences, a single result may not universally satisfy diverse user tastes. A natural

extension is to implement test-time scaling by generating a spectrum of candidates with varying styles

and compositions. This paradigm shifts the focus from single-shot generation to a generate-and-select

workﬂow, enabling either human users or VLM-based preference models to select the illustration that

best aligns with their speciﬁc requirements.

Extension to Broader Domains. Beyond academic illustrations, our framework establishes a gener-

alizable paradigm: leveraging retrieval to instruct the model on what to generate (target diagram

types) and employing automatic style summarization to teach it how to generate (stylistic norms). By

eﬀectively decoupling structural planning from aesthetic rendering, this reference-driven approach

bypasses the need for expensive domain-speciﬁc ﬁne-tuning. We believe this paradigm holds signiﬁ-

cant promise for other specialized domains requiring strict adherence to community standards, such

as UI/UX design, patent drafting, and industrial schematics.

Acknowledgements

We thank all members of Google Cloud AI Research for their valuable support during the project. We

also thank Yuhang and Ali for the thoughtful discussion.

PaperBanana: Automating Academic Illustration for AI Scientists

Impact Statement

This paper introduces PaperBanana, a framework designed to automate the generation of academic

illustrations. Our goal is to democratize access to high-quality visual communication tools, particularly

beneﬁting researchers who may lack professional design resources. By reducing the manual eﬀort

required for diagram creation, we aim to accelerate the scientiﬁc workﬂow. However, we acknowledge

the ethical risk associated with generative models, speciﬁcally the potential for “visual hallucination”

or unfaithful representation of technical details. It is imperative that users of such systems reject

blind reliance and maintain rigorous human oversight to ensure the scientiﬁc integrity of published

illustrations.

References

Anonymous. Autoﬁgure: Generating and reﬁning publication-ready scientiﬁc illustrations. In Submitted

to International Conference on Learning Representations, 2026. URL

https://openreview.net/

pdf?id=5N3z9JQJKq. under review.

Anthropic. Claude sonnet 4: Hybrid reasoning model with superior intelligence for high-volume use

cases, and 200k context window. https://www.anthropic.com/claude/sonnet, 2025.

J. Belouadi and S. Eger. Detikzify: Synthesizing graphics programs for scientiﬁc ﬁgures and sketches

with tikz. In Advances in Neural Information Processing Systems, 2024.

J. Belouadi, E. Ilg, M. Keuper, H. Tanaka, M. Utiyama, R. Dabre, S. Eger, and S. Ponzetto. Tikzero:

Zero-shot text-guided graphics program synthesis. In Proceedings of the IEEE/CVF International

Conference on Computer Vision, pages 17793–17806, 2025.

BIT-DataLab. Edit banana, Oct. 2025. URL

https://github.com/BIT-DataLab/Edit-Banana

Y. Chang, Y. Feng, J. Sun, J. Ai, C. Li, S. K. Zhou, and K. Zhang. Sridbench: Benchmark of scientiﬁc

research illustration drawing of image generation model. arXiv preprint arXiv:2505.22126, 2025.

Z. Chen, J. Chen, S. Ö. Arik, M. Sra, T. Pﬁster, and J. Yoon. Coda: Agentic systems for collaborative

data visualization. arXiv preprint arXiv:2510.03194, 2025.

J. Cohen. Statistical power analysis for the behavioral sciences. routledge, 2013.

G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram,

D. Zhang, E. Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality,

long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025.

G. Deepmind. Introducing nano banana pro.

https://blog.google/technology/ai/

nano-banana-pro/, 2025.

V. Dibia. Lida: A tool for automatic generation of grammar-agnostic visualizations and infographics

using large language models. In Proceedings of the 61st Annual Meeting of the Association for

Computational Linguistics (Volume 3: System Demonstrations), pages 113–126, 2023.

V. Dibia and Ç. Demiralp. Data2vis: Automatic generation of data visualizations using sequence-

to-sequence recurrent neural networks. IEEE computer graphics and applications, 39(5):33–46,

2019.

PaperBanana: Automating Academic Illustration for AI Scientists

T. Galimzyanov, S. Titov, Y. Golubev, and E. Bogomolov. Drawing pandas: A benchmark for llms in

generating plotting code. In 2025 IEEE/ACM 22nd International Conference on Mining Software

Repositories (MSR), pages 503–507. IEEE, 2025.

Z. Ghahramani. Probabilistic machine learning and artiﬁcial intelligence. Nature, 521(7553):452–459,

2015.

K. Goswami, P. Mathur, R. Rossi, and F. Dernoncourt. Plotgen: Multi-agent llm-based scientiﬁc data

visualization via multimodal feedback. arXiv preprint arXiv:2502.00988, 2025.

J. Gottweis, W.-H. Weng, A. Daryin, T. Tu, A. Palepu, P. Sirkovic, A. Myaskovsky, F. Weissenberger,

K. Rong, R. Tanno, et al. Towards an ai co-scientist. arXiv preprint arXiv:2502.18864, 2025.

M. Hollander, D. A. Wolfe, and E. Chicken. Nonparametric statistical methods. John Wiley & Sons,

2013.

W. K. A. L. J. B. A. Hsu and S. Eger. Automatikz: Text-guided synthesis of scientiﬁc vector graphics

with tikz. arXiv preprint arXiv:2310.00367, 2023.

S. Huang, Y. Gao, J. Bai, Y. Zhou, Z. Yin, X. Liu, R. Chellappa, C. P. Lau, S. Nag, C. Peng, et al. Sciﬁg:

Towards automating scientiﬁc ﬁgure generation. arXiv preprint arXiv:2601.04390, 2026.

P. Langley. Scientiﬁc discovery: Computational explorations of the creative processes. MIT press, 1987.

P. Langley. Integrated systems for computational scientiﬁc discovery. In Proceedings of the AAAI

Conference on Artiﬁcial Intelligence, volume 38, pages 22598–22606, 2024.

S. Li, X. Chen, Y. Song, Y. Song, and C. Zhang. Prompt4vis: Prompting large language models with

example mining and schema ﬁltering for tabular data visualization. arXiv preprint arXiv:2402.07909,

2024.

S. Li, Y. Zhang, J. Wu, Z. Lei, Y. He, R. Wen, C. Liao, C. Jiang, A. Ping, S. Gao, et al. If-vidcap: Can

video caption models follow instructions? arXiv preprint arXiv:2510.18726, 2025.

C. Liang and J. You. Diagrameval: Evaluating llm-generated diagrams via graphs. arXiv preprint

arXiv:2510.25761, 3, 2025.

A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruan, et al. Deepseek-v3

technical report. arXiv preprint arXiv:2412.19437, 2024.

Z. Liu, Z. Guo, Z. Su, L. Huang, Y. Yang, Z. Han, Z. Pan, and W. Zhang. Paper2Any: Turn Paper/Tex-

t/Topic into Editable Research Figures, Technical Route Diagrams, and Presentation Slides, Oct.

2025. URL https://github.com/OpenDCAI/Paper2Any.

C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune, and D. Ha. The ai scientist: Towards fully automated

open-ended scientiﬁc discovery. arXiv preprint arXiv:2408.06292, 2024.

Z. Luo, Z. Yang, Z. Xu, W. Yang, and X. Du. Llm4sr: A survey on large language models for scientiﬁc

research. arXiv preprint arXiv:2501.04306, 2025.

J. Mackinlay. Automating the design of graphical presentations of relational information. Acm

Transactions On Graphics (Tog), 5(2):110–141, 1986.

PaperBanana: Automating Academic Illustration for AI Scientists

J. Niu, Z. Liu, Z. Gu, B. Wang, L. Ouyang, Z. Zhao, T. Chu, T. He, F. Wu, Q. Zhang, Z. Jin, G. Liang,

R. Zhang, W. Zhang, Y. Qu, Z. Ren, Y. Sun, Y. Zheng, D. Ma, Z. Tang, B. Niu, Z. Miao, H. Dong,

S. Qian, J. Zhang, J. Chen, F. Wang, X. Zhao, L. Wei, W. Li, S. Wang, R. Xu, Y. Cao, L. Chen, Q. Wu,

H. Gu, L. Lu, K. Wang, D. Lin, G. Shen, X. Zhou, L. Zhang, Y. Zang, X. Dong, J. Wang, B. Zhang,

L. Bai, P. Chu, W. Li, J. Wu, L. Wu, Z. Li, G. Wang, Z. Tu, C. Xu, K. Chen, Y. Qiao, B. Zhou, D. Lin,

W. Zhang, and C. He. Mineru2.5: A decoupled vision-language model for eﬃcient high-resolution

document parsing, 2025. URL https://arxiv.org/abs/2509.22186.

OpenAI. Introducing chatgpt. 2022. URL https://openai.com/blog/chatgpt.

OpenAI. Gpt-image-1. https://platform.openai.com/docs/models/gpt-image-1, 2025a.

OpenAI. Introducing gpt-5. https://openai.com/index/introducing-gpt-5/, 2025b.

W. Pang, K. Q. Lin, X. Jian, X. He, and P. Torr. Paper2poster: Towards multimodal poster automation

from scientiﬁc papers. arXiv preprint arXiv:2505.21497, 2025.

A. Quispel, A. Maes, and J. Schilperoord. Aesthetics and clarity in information visualization: The

designer’s perspective. In Arts, volume 7, page 72. MDPI, 2018.

J. A. Rodriguez, D. Vazquez, I. Laradji, M. Pedersoli, and P. Rodriguez. Figgen: Text to scientiﬁc ﬁgure

generation. arXiv preprint arXiv:2306.00800, 2023.

J. Schmidhuber. Artiﬁcial scientists & artists based on the formal theory of creativity. In 3d Conference

on Artiﬁcial General Intelligence (AGI-2010), pages 148–153. Atlantis Press, 2010.

W. Seo, S. Lee, D. Kang, H. An, Z. Yuan, and S. Lee. Automated visualization code synthesis via

multi-path reasoning and feedback-driven optimization. arXiv preprint arXiv:2502.11140, 2025.

N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao. Reﬂexion: Language agents with verbal

reinforcement learning. Advances in Neural Information Processing Systems, 36:8634–8652, 2023.

C. Snell, J. Lee, K. Xu, and A. Kumar. Scaling llm test-time compute optimally can be more eﬀective

than scaling model parameters. arXiv preprint arXiv:2408.03314, 2024.

J. Sun, F. Zhang, Y. Feng, C. Li, Z. Li, J. Ai, Y. Chang, Y. Dai, and K. Zhang. From pixels to paths: A

multi-agent framework for editable scientiﬁc illustration. arXiv preprint arXiv:2510.27452, 2025.

Y. Tang, X. Liu, B. Zhang, T. Lan, Y. Xie, J. Lao, Y. Wang, H. Li, T. Gao, B. Pan, et al. Igenbench:

Benchmarking the reliability of text-to-infographic generation. arXiv preprint arXiv:2601.04498,

2026.

K. Team, J. Chen, Y. Ci, X. Du, Z. Feng, K. Gai, S. Guo, F. Han, J. He, K. He, et al. Kling-omni technical

report. arXiv preprint arXiv:2512.16776, 2025.

Y. Tian, W. Cui, D. Deng, X. Yi, Y. Yang, H. Zhang, and Y. Wu. Chartgpt: Leveraging llms to generate

charts from abstract natural language. IEEE Transactions on Visualization and Computer Graphics,

31(3):1731–1745, 2024.

E. R. Tufte and P. R. Graves-Morris. The visual display of quantitative information, volume 2. Graphics

press Cheshire, CT, 1983.

C. Wu, J. Li, J. Zhou, J. Lin, K. Gao, K. Yan, S.-m. Yin, S. Bai, X. Xu, Y. Chen, et al. Qwen-image

technical report. arXiv preprint arXiv:2508.02324, 2025a.

PaperBanana: Automating Academic Illustration for AI Scientists

C. Wu, Z. Liang, Y. Ge, Q. Guo, Z. Lu, J. Wang, Y. Shan, and P. Luo. Plot2code: A comprehensive

benchmark for evaluating multi-modal large language models in code generation from scientiﬁc

plots. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 3006–3028,

2025b.

A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, et al. Qwen3

technical report. arXiv preprint arXiv:2505.09388, 2025a.

C. Yang, C. Shi, Y. Liu, B. Shui, J. Wang, M. Jing, L. XU, X. Zhu, S. Li, Y. Zhang, et al. Chartmimic:

Evaluating lmm’s cross-modal reasoning capability via chart-to-code generation. In The Thirteenth

International Conference on Learning Representations, 2025b.

Z. Yang, Z. Zhou, S. Wang, X. Cong, X. Han, Y. Yan, Z. Liu, Z. Tan, P. Liu, D. Yu, et al. Matplotagent:

Method and evaluation for llm-based agentic scientiﬁc data visualization. In Findings of the

Association for Computational Linguistics ACL 2024, pages 11789–11804, 2024.

L. Zhang, S. Eger, Y. Cheng, W. ZHAI, J. Belouadi, F. Moaﬁan, and Z. Zhao. Scimage: How good

are multimodal large language models at scientiﬁc text-to-image generation? In The Thirteenth

International Conference on Learning Representations, 2025.

H. Zheng, X. Guan, H. Kong, W. Zhang, J. Zheng, W. Zhou, H. Lin, Y. Lu, X. Han, and L. Sun. Pptagent:

Generating and evaluating presentations beyond text-to-slides. In Proceedings of the 2025 Conference

on Empirical Methods in Natural Language Processing, pages 14413–14429, 2025.

J. Zuo, H. Deng, H. Zhou, J. Zhu, Y. Zhang, Y. Zhang, Y. Yan, K. Huang, W. Chen, Y. Deng, et al. Is

nano banana pro a low-level vision all-rounder? a comprehensive evaluation on 14 tasks and 40

datasets. arXiv preprint arXiv:2512.15110, 2025.

PaperBanana: Automating Academic Illustration for AI Scientists

A. Dedicated Case Studies

Cases Demonstrating the Eﬀectiveness of PaperBanana We provide 2 cases in Figure 7to

demonstrate the capability of PaperBanana for aiding the generation of academic illustrations.

Given the same source context and caption, the vanilla Nano-Banana-Pro often produces diagrams

with outdated color tones and overly verbose content. In contrast, our PaperBanana generates

results that are more concise and aesthetically pleasing, while maintaining faithfulness to the source

context.

Case 1

Case 2

Human

Nano-

Banana

-Pro

🍌

Paper-

Banana

Analy-

sis

In general, while being faithful to the source context, PaperBanana demonstrates clear advantages over vanilla

Nano-Banana-Pro across multiple dimensions.

1. From an aesthetic perspective, Case 1 and Case 2 both show that PaperBanana employs more modern and

harmonious color palettes (e.g., soft blue and orange zones) compared to the outdated grey tones in Nano-Banana-Pro.

2. Regarding conciseness, PaperBanana effectively distills core information while eliminating redundant textual

descriptions, making the diagrams cleaner and more focused.

3. In terms of logical clarity, PaperBanana utilizes distinct colored blocks and zones to separate different functional

modules (e.g., ``Lens Module'' and ``Reasoning Module'' in Case 1; ``Confidence Estimation (ECE)'' and ``Volume \&

Disparity'' in Case 2), thereby enhancing the readability and comprehensibility of complex methodological workflows.

Figure 7

Case study of diagram generation. Given the same source context and caption, the vanilla

Nano-Banana-Pro often produces diagrams with outdated color tones and overly verbose content. In

contrast, our PaperBanana generates results that are more concise and aesthetically pleasing, while

maintaining faithfulness to the source context.

PaperBanana: Automating Academic Illustration for AI Scientists

Enhancing the Aesthetics of Human-Drawn Diagrams We provide additional cases in Figure 8to

demonstrate the interesting scenario of enhancing the aesthetics of human-drawn diagrams with our

auto-summarized style guidelines. It is observed that the polished diagrams demonstrate signiﬁcant

stylistic improvements in color schemes, typography, graphical elements, etc.

Original Diagrams

Style Enhanced Diagrams

Figure 8

Additional cases for enhancing the aesthetics of human-drawn diagrams with our auto-

summarized style guidelines. The polished diagrams demonstrate signiﬁcant stylistic improvements

in color schemes, typography, graphical elements, etc.

Case study for visualizing statistical plots with code and image generation. Figure 9compares

the results of visualizing statistical plots with code and image generation. It is observed that the

image generation model can generate more visually appealing plots, but incurs more faithfulness

errors such as numerical hallucination or element repetition.

PaperBanana: Automating Academic Illustration for AI Scientists

Plots Visualized via IMG

Plots Visualized via Coding

Case Analysis

Both plots are correct. The left

plot looks more visually

appealing

Both plots are correct. The left

plot looks more visually

appealing

The left plot looks better, but

contains faithfulness issues:

On the `Rebound` axis, it

depicts Player A (Blue) as

having a significantly higher

value (~0.9) than Player B

(Orange, ~0.6), effectively

inverting the relationship and

plotting the wrong values

(likely duplicating the 'Steals'

data).

The left plot contains

faithfulness issues:

It duplicates the “East 10”

category.

The left plot contains

faithfulness errors:

The 'Clinical' data value is 0.4,

but the Model draws the bar

significantly taller than the 0.4

gridline and axis tick

Figure 9

Case study for visualizing statistical plots with code and image generation. It is observed that

the image generation model can generate more visually appealing plots, but incurs more faithfulness

errors such as numerical hallucination or element repetition. The red bounding boxes are added by

the authors to highlight the errors.

PaperBanana: Automating Academic Illustration for AI Scientists

Failure Cases of PaperBanana.Figure 10 shows 3 failure cases of PaperBanana. We observe that

the primary failure mode involves connection errors, such as redundant connections and mismatched

source-target nodes. Our preliminary analysis reveals that the critic model often fails to identify

these connectivity issues, suggesting these errors may originate from the foundation model’s inherent

perception limitations. Resolving this challenge likely necessitates advancements in the underlying

foundation model.

Diagrams from PaperBanana

Failure Reasons

The attention block incorrectly draws

connections from the Noise Predictor (

) to three destination boxes,

𝑓𝑠𝑑

labeling two of them as 'Keys ( )'

𝐾𝑖

and one as 'Values ( )'. The text (Eq.

𝑉𝑖

2) explicitly states that diffusion

features map only to Keys and Values,

while Queries ( ) come from

𝑄𝑖

learnable parameters.

It fails to draw generative connections

(solid lines) from Actions ( ) to

𝑎

Task-Relevant States ( ). The method 𝑠

explicitly states (Eq. 3) that state

transitions depend on previous actions

( ). By isolating the

ℎ'𝑡=𝑓(...,𝑎𝑡−1)

actions without driving the state

dynamics, the diagram fails to

represent the fundamental structure of

a control/world model described in the

text.

The text explicitly states, "We note

that the Decoder and Predictor can

only access information of t through

e" and defines the Decoder's input as

the embedding and the goal (

D(e,g))). However, the Model diagram

draws a dotted skip connection from

the "Tactic t" block directly to the

Decoder, replacing the required

"Goal" connection. This violates the

functional constraints described in the

methodology.

Figure 10

Failure cases of PaperBanana. The primary failure mode involves connection errors,

such as redundant connections and mismatched source-target nodes. Our preliminary analysis reveals

that the critic model often fails to identify these connectivity issues, suggesting these errors may

originate from the foundation model’s inherent perception limitations. Resolving this challenge likely

necessitates advancements in the underlying foundation model.

PaperBanana: Automating Academic Illustration for AI Scientists

B. Human Evaluation Setup

To ensure the reliability of our automated metrics and strict benchmarking of our method, this

paper conducted two distinct human evaluation experiments. Both evaluations employed the same

four dimensions deﬁned in Section 4(Faithfulness, Conciseness, Readability, and Aesthetics) and

adhered to the same detailed rubrics used by our VLM judge. We utilized Streamlit to build dedicated

annotation interfaces for these tasks.

Validation of VLM-as-a-Judge. The objective of this human evaluation is to assess the alignment

between our VLM-based judge (Gemini-3-Pro) and human judgment. We randomly sampled 50

cases (25 from the Vanilla baseline and 25 from PaperBanana) from the test set. For each case,

two experienced researchers were presented with the Method Section, Caption, the human-drawn

reference diagram, and a model-generated candidate (either from our method or the baseline). They

were tasked with conducting a side-by-side comparison on the four evaluation dimensions. For

conﬂicting cases, they engaged in discussion to reach a consensus. For each dimension, the annotator

selected one of four outcomes: “Model wins”, “Human wins”, “Both are good”, or “Both are bad”.

These choices were then mapped to numerical scores (100, 0, 50, 50) to calculate the Kendall’s tau

correlation with the VLM judge’s scores, as reported in Section 5. The annotation interface is shown

in Figure 11.

Blind Test for Main Results. To rigorously compare PaperBanana against the strong baseline

(Vanilla Nano-Banana-Pro), we conducted a blind A/B test on a subset of 50 cases. Three experienced

researchers were presented with the Method Section, Caption, a Reference (Human Drawn) diagram,

and two anonymous candidates (Candidate A and Candidate B) in randomized order. To determine

the winner, we enforced a hierarchical decision strategy consistent with our VLM evaluation protocol.

Annotators ﬁrst evaluated the Primary Dimensions (Faithfulness and Readability). If a candidate won

in the primary dimensions (or won one and tied the other), it was declared the overall winner. In cases

of a tie in primary dimensions, the decision was deferred to the Secondary Dimensions (Conciseness

and Aesthetics). This setup ensures that our human evaluation prioritizes content correctness and

clarity, mirroring the rigorous standards of academic publication. The annotation interface is shown

in Figure 12.

Figure 11 |Annotation interface for reference-based evaluation.

PaperBanana: Automating Academic Illustration for AI Scientists

Figure 12 |Annotation interface for blind human evaluation.

C. Implementation Details

Categorization of Methodology Diagrams. To facilitate detailed analysis, we categorize the

diagrams into four classes based on visual topology and content. The detailed deﬁnitions and

keywords for each category are listed in Table 3.

Additional Experiment Settings. For all experiments, we set the generation temperature to 1. To

ensure fair comparisons, we align the aspect ratio of the generated images with their human-drawn

counterparts. Speciﬁcally, we calculate the aspect ratio of the ground-truth diagram and match it to

the nearest ratio supported by the image generation model (e.g., for Nano-Banana-Pro, we round to

the closest among 3:2, 16:9, and 21:9).

Generating Diagrams and Plots used in this Paper. All ﬁgures in this paper marked with “[Gener-

ated by ]” are produced entirely by PaperBanana. In practice, given the inherent variability of

generative models, we generated multiple candidates and manually selected the best one for presenta-

tion. We recommend this “generate-and-select” workﬂow for practical applications of PaperBanana.

D. Testset Curation for Statistical Plots Generation

This section introduces the testset curation process for statistical plots generation, which evaluates

the capability to generate statistical plots from raw data (e.g., tables, CSV ﬁles) and high-level visual

descriptions (e.g. a bar plot titled "Number of Publications by Year"). Since academic manuscripts

rarely include raw data for their published plots, we repurpose ChartMimic (Yang et al.,2025b), a

dataset originally designed for chart-to-code evaluation. Speciﬁcally, we use the “direct mimic” subset,

which contains 2,400 plots sourced majorly from arXiv papers and matplotlib galleries, each paired

with human-curated Python code for reproduction. This enables us to systematically extract both the

underlying data and visual descriptions, while using the plots themselves as ground truth. Speciﬁcally,

the pipeline is as follows:

PaperBanana: Automating Academic Illustration for AI Scientists

Table 3 |Categorization of diagrams based on visual topology and content.

1. Agent & Reasoning

•LLM agents, multi-agent systems, reasoning, planning, tool use

•Instruction following, in-context learning, chain-of-thought

•Code generation, autonomous systems

•Keywords: agent, llm, language model, reasoning, planning, prompt

2. Vision & Perception

•Computer vision, 3D reconstruction, rendering, object detection

•Scene understanding, depth estimation, pose estimation

•Visual representations and feature learning

•Keywords: vision, image, 3d, gaussian, nerf, detection, segmentation, camera

3. Generative & Learning

•Generative models (diﬀusion, GANs, VAEs, autoencoders)

•Reinforcement learning, policy learning

•Optimization and training dynamics

•Keywords: diﬀusion, generative, gan, denoising, reinforcement, policy, reward

4. Science & Applications

•AI for Science (biology, chemistry, physics, medicine)

•Graph neural networks, structured data

•Theoretical analysis, mathematical proofs

•Domain-speciﬁc applications

•Keywords: protein, molecule, biology, graph, node, theorem, theory

Collection & Filtering. We begin with all 2,400 plots from the “direct mimic” subset. Using Gemini-

3-Pro, we extract the raw data from the code into tabular format, generate a high-level description of

each plot’s visual intent, while also marking the diﬃculty of generating the plot (Speciﬁcally, plots

with many data points or subplots are marked as diﬃcult, while plots with only 1 subplot and few data

points are marked as easy). Meanwhile, we also apply two ﬁltering criteria: (1) Reproducible Data:

exclude plots where data is randomly generated or requires complex computations; (2) Standard

Mapping: exclude plots using data for geometric construction (e.g., drawing shapes) rather than

conventional statistical visualization. Similar to our methodology diagram curation, we ﬁlter out

plots with aspect ratios (

𝑤

ℎ

) outside [1.0, 2.5] to support future exploration with image generation

models. This yields 914 plots.

Categorization. ChartMimic’s original 22 plot categories include many types rarely used in academic

publications, such as Pip chart and Quiver chart. Based on the distribution of our 914 ﬁltered plots,

we consolidate them into 7 common categories: Bar Chart,Line Chart,Tree & Pie Chart,Scatter Plot,

Heatmap,Radar Chart, and Miscellaneous (all other types).

Sampling. We then sample 80 plots per category, except for Heatmap and Radar Chart (40 each

due to limited availability), yielding 480 plots total. During sampling, we intentionally increased the

proportion of diﬃcult cases to ensure a challenging testset. Each category is then evenly split into

reference and test sets.

PaperBanana: Automating Academic Illustration for AI Scientists

E. Textual Description of our Methodology Diagram

Our framework operates by ﬁrst synthesizing a detailed description of the target diagram, which

is then visualized by Nano-Banana-Pro. To facilitate reproduction and inspire future research, we

provide below the exact textual description synthesized by our framework during the actual inference

run that produced Figure 2, which served as the input to the Visualizer. When using Nano-Banana-Pro,

we set the (width:height) aspect ratio as 21:9, temperature as 1, and resolution as 2K.

Textual Description of our Methodology Diagram

The figure is a wide, horizontal flowchart-style diagram illustrating the "

Paperbanana" framework. The layout flows from left to right on a clean

white background, divided into two main colored regions: the "Linear

Planning Phase" (left/middle) and the "Iterative Refinement Loop" (right)

**1. Leftmost Section: Inputs**

* **Visual Elements:** Two icons stacked vertically on the far left.

* Top: A document icon labeled **"Source Context ($S$)"**.

* Bottom: A target/goal icon labeled **"Communicative Intent ($C$)"**.

* **Flow:** Brackets merge these inputs into a main flow line that enters

the first phase.

**2. Middle-Left Region: Linear Planning Phase**

* **Container:** A light blue rounded rectangle. Label at top: **"Linear

Planning Phase"**.

* **Reference Set ($\mathcal{R}$):** A cylinder database icon located at

the bottom-left of this region, labeled **"Reference Set ($\mathcal{R}$)

"**.

* **Agent 1: Retriever Agent**

* **Icon:** Robot with a magnifying glass.

* **Label:** **"Retriever Agent"** positioned below the icon.

* **Input:** An arrow from the main Inputs ($S, C$) and an arrow from

the Reference Set ($\mathcal{R}$).

* **Output:** Arrow to a cluster of image thumbnails labeled **"

Relevant Examples ($\mathcal{E}$)"**.

* **Agent 2: Planner Agent**

* **Icon:** Robot with a clipboard or thought bubble.

* **Label:** **"Planner Agent"** positioned below the icon.

* **Input:** Receives an arrow from "Relevant Examples ($\mathcal{E}$)

". **Crucially**, a direct flow arrow (bypassing the Retriever) connects

the main Inputs ($S, C$) to the Planner, indicating it uses the source

content for planning.

* **Output:** Arrow to a text document icon labeled **"Initial

Description ($P$)"**.

* **Agent 3: Stylist Agent**

* **Icon:** Robot with a palette/paintbrush.

* **Label:** **"Stylist Agent"** positioned below the icon.

* **Input:** Receives "Initial Description ($P$)" and a dashed arrow

from the Reference Set ($\mathcal{R}$) labeled **"Aesthetic Guidelines ($

\mathcal{G}$)"**.

* **Output:** An arrow exiting the blue region labeled **"Optimized

Description ($P^*)"**.

PaperBanana: Automating Academic Illustration for AI Scientists

**3. Middle-Right Region: Iterative Refinement Loop**

* **Container:** A light orange rounded rectangle. Label at top: **"

Iterative Refinement Loop"**.

* **Agent 4: Visualizer Agent**

* **Icon:** Robot standing next to a split visual representation: a

canvas on one side and a code terminal/brackets (‘</>‘) on the other.

* **Label:** **"Visualizer Agent"** positioned below the icon.

* **Input:** Takes "Optimized Description ($P^*)" (from Stylist) and "

Refined Description ($P_{t+1}$)" (from Critic).

* **Output:** Arrow to an image preview labeled **"Generated Image (

$I_t$)"**.

* **Agent 5: Critic Agent**

* **Icon:** Robot with a checklist/reviewer pen.

* **Label:** **"Critic Agent"** positioned below the icon.

* **Input:** Receives "Generated Image ($I_t$)". A long **dashed gray

line** labeled **"Factual Verification"** runs from the original Inputs (

$S, C$) along the bottom edge, connecting to the Critic.

* **Output:** A curved return arrow back to the Visualizer, labeled

**"Refined Description ($P_{t+1}$)"**.

* **Center Element:** A circular arrow icon inside the loop indicating **"

$T=3$ Rounds"**.

**4. Rightmost Section: Final Output**

* **Visual Element:** A polished scientific illustration emerging from the

loop.

* **Label:** **"Final Illustration ($I_T$)"**.

**5. Styling**

* **Agents:** Cute, consistent robot avatars with distinct accessories.

* **Typography:** Sans-serif for main text. **Serif Italic (LaTeX style)**

for all variables ($S, C, P, I, \mathcal{R}, \mathcal{E}, \mathcal{G}$).

* **Colors:** Blue accents for Planning; Orange accents for Refinement.

Main flow arrows in solid black; secondary inputs in dashed gray.

F. Auto Summarized Style Guide for Academic Illustrations

F.1. Style Guides for Methodology Diagrams and Statistical Plots

Style Guide for Methodology Diagrams

### 1. The "NeurIPS Look"

The prevailing aesthetic for 2025 is **"Soft Tech & Scientific Pastels."**

Gone are the days of harsh primary colors and sharp black boxes. The modern

NeurIPS diagram feels approachable yet precise. It utilizes high-value (

light) backgrounds to organize complexity, reserving saturation for the

most critical active elements. The vibe balances **clean modularity** (

clear separation of parts) with **narrative flow** (clear left-to-right

progression).

---

### 2. Detailed Style Options

PaperBanana: Automating Academic Illustration for AI Scientists

#### **A. Color Palettes**

*Design Philosophy: Use color to group logic, not just to decorate. Avoid

fully saturated backgrounds.*

**Background Fills (The "Zone" Strategy)**

*Used to encapsulate stages (e.g., "Pre-training phase") or environments.*

* **Most papers use:** Very light, desaturated pastels (Opacity ~10-15%).

* **Aesthetically pleasing options include:**

* **Cream / Beige** (e.g., ‘#F5F5DC‘) - *Warm, academic feel.*

* **Pale Blue / Ice** (e.g., ‘#E6F3FF‘) - *Clean, technical feel.*

* **Mint / Sage** (e.g., ‘#E0F2F1‘) - *Soft, organic feel.*

* **Pale Lavender** (e.g., ‘#F3E5F5‘) - *distinctive, modern feel.*

* **Alternative (~20%):** White backgrounds with colored *dashed borders*

for a high-contrast, minimalist look (common in theoretical papers).

**Functional Element Colors**

* **For "Active" Modules (Encoders, MLP, Attention):** Medium saturation

is preferred.

* *Common pairings:* Blue/Orange, Green/Purple, or Teal/Pink.

* *Observation:* Colors are often used to distinguish **status**

rather than component type:

* **Trainable Elements:** Often Warm tones (Red, Orange, Deep Pink

* **Frozen/Static Elements:** Often Cool tones (Grey, Ice Blue,

Cyan).

* **For Highlights/Results:** High saturation (Primary Red, Bright Gold)

is strictly reserved for "Error/Loss," "Ground Truth," or the final

output.

#### **B. Shapes & Containers**

*Design Philosophy: "Softened Geometry." Sharp corners are for data; rounded

corners are for processes.*

**Core Components**

* **Process Nodes (The Standard):** Rounded Rectangles (Corner radius 5-10

px). This is the dominant shape (~80%) for generic layers or steps.

* **Tensors & Data:**

* **3D Stacks/Cuboids:** Used to imply depth/volume (e.g., $B \times H

\times W$).

* **Flat Squares/Grids:** Used for matrices, tokens, or attention maps.

* **Cylinders:** Exclusively reserved for Databases, Buffers, or

Memory.

**Grouping & Hierarchy**

* **The "Macro-Micro" Pattern:** A solid, light-colored container

represents the global view, with a specific module (e.g., "Attention

Block") connected via lines to a "zoomed-in" detailed breakout box.

* **Borders:**

* **Solid:** For physical components.

* **Dashed:** Highly prevalent for indicating "Logical Stages," "

Optional Paths," or "Scopes."

PaperBanana: Automating Academic Illustration for AI Scientists

#### **C. Lines & Arrows**

*Design Philosophy: Line style dictates flow type.*

**Connector Styles**

* **Orthogonal / Elbow (Right Angles):** Most papers use this for **

Network Architectures** (implies precision, matrices, and tensors).

* **Curved / Bezier:** Common choices include this for **System Logic,

Feedback Loops, or High-Level Data Flow** (implies narrative and

connection).

**Line Semantics**

* **Solid Black/Grey:** Standard data flow (Forward pass).

* **Dashed Lines:** Universally recognized as "Auxiliary Flow."

* *Used for:* Gradient updates, Skip connections, or Loss calculations.

* **Integrated Math:** Standard operators ($\oplus$ for Add, $\otimes$ for

Concat/Multiply) are frequently placed *directly* on the line or

intersection.

#### **D. Typography & Icons**

*Design Philosophy: Strict separation between "Labeling" and "Math."*

**Typography**

* **Labels (Module Names):** **Sans-Serif** (Arial, Roboto, Helvetica).

* *Style:* Bold for headers, Regular for details.

* **Variables (Math):** **Serif** (Times New Roman, LaTeX default).

* *Rule:* If it is a variable in your equation (e.g., $x, \theta, \

mathcal{L}$), it **must** be Serif and Italicized in the diagram.

**Iconography Options**

* **For Model State:**

* *Trainable:* Fire, Lightning.

* *Frozen:* Snowflake, Padlock, Stop Sign (Greyed out).

* **For Operations:**

* *Inspection:* Magnifying Glass.

* *Processing/Computation:* Gear, Monitor.

* **For Content:**

* *Text/Prompt:* Document, Chat Bubble.

* *Image:* Actual thumbnail of an image (not just a square).

---

### 3. Common Pitfalls (How to look "Amateur")

* **The "PowerPoint Default" Look:** Using standard Blue/Orange presets

with heavy black outlines.

* **Font Mixing:** Using Times New Roman for "Encoder" labels (makes the

paper look dated to the 1990s).

* **Inconsistent Dimension:** Mixing flat 2D boxes and 3D isometric cubes

without a clear reason (e.g., 2D for logic, 3D for tensors is fine;

random mixing is not).

* **Primary Backgrounds:** Using saturated Yellow or Blue backgrounds for

grouping (distracts from the content).

* **Ambiguous Arrows:** Using the same line style for "Data Flow" and "

Gradient Flow."

PaperBanana: Automating Academic Illustration for AI Scientists

---

### 4. Domain-Specific Styles

**If you are writing an AGENT / LLM Paper:**

* **Vibe:** Illustrative, Narrative, "Friendly.", Cartoony.

* **Key Elements:** Use "User Interface" aesthetics. Chat bubbles for

prompts, document icons for retrieval.

* **Characters:** It is common to use cute 2D vector robots, human avatars,

or emojis to humanize the agent’s reasoning steps.

**If you are writing a COMPUTER VISION / 3D Paper:**

* **Vibe:** Spatial, Dense, Geometric.

* **Key Elements:** Frustums (camera cones), Ray lines, and Point Clouds.

* **Color:** Often uses RGB color coding to denote axes or channel

correspondence. Use heatmaps (Rainbow/Viridis) to show activation.

**If you are writing a THEORETICAL / OPTIMIZATION Paper:**

* **Vibe:** Minimalist, Abstract, "Textbook."

* **Key Elements:** Focus on graph nodes (circles) and manifolds (planes/

surfaces).

* **Color:** Restrained. mostly Grayscale/Black/White with one highlight

color (e.g., Gold or Blue). Avoid "cartoony" elements.

Style Guide for Statistical Plots

# NeurIPS 2025 Statistical Plot Aesthetics Guide

## 1. The "NeurIPS Look": A High-Level Overview

The prevailing aesthetic for 2025 is defined by **precision, accessibility,

and high contrast**. The "default" academic look has shifted away from

bare-bones styling toward a more graphic, publication-ready presentation.

* **Vibe:** Professional, clean, and information-dense.

* **Backgrounds:** There is a heavy bias toward **stark white backgrounds

** for maximum contrast in print and PDF reading, though the "Seaborn-

style" light grey background remains an accepted variant.

* **Accessibility:** A strong emphasis on distinguishing data not just by

color, but by texture (patterns) and shape (markers) to support black-and

-white printing and colorblind readers.

---

## 2. Detailed Style Options

### **Color Palettes**

* **Categorical Data:**

* **Soft Pastels:** Matte, low-saturation colors (salmon, sky blue,

mint, lavender) are frequently used to prevent visual fatigue.

* **Muted Earth Tones:** "Academic" palettes using olive, beige, slate

grey, and navy.

PaperBanana: Automating Academic Illustration for AI Scientists

* **High-Contrast Primaries:** Used sparingly when categories must be

distinct (e.g., deep orange vs. vivid purple).

* **Accessibility Mode:** A growing trend involves combining color

with **geometric patterns** (hatches, dots, stripes) to differentiate

categories.

* **Sequential & Heatmaps:**

* **Perceptually Uniform:** "Viridis" (blue-to-yellow) and "Magma/

Plasma" (purple-to-orange) are the standard.

* **Diverging:** "Coolwarm" (blue-to-red) is used for positive/

negative value splits.

* **Avoid:** The traditional "Jet/Rainbow" scale is almost entirely

absent.

### **Axes & Grids**

* **Grid Style:**

* **Visibility:** Grid lines are almost rarely solid. Common choices

include **fine dashed (‘--‘)** or **dotted (‘:‘)** lines in light gray.

* **Placement:** Grids are consistently rendered *behind* data

elements (low Z-order).

* **Spines (Borders):**

* **The "Boxed" Look:** A full enclosure (black spines on all 4 sides)

is very common.

* **The "Open" Look:** Removing the top and right spines for a

minimalist appearance.

* **Ticks:**

* **Style:** Ticks are generally subtle, facing inward, or removed

entirely in favor of grid alignment.

### **Layout & Typography**

* **Typography:**

* **Font Family:** Exclusively **Sans-Serif** (resembling Helvetica,

Arial, or DejaVu Sans). Serif fonts are rarely used for labels.

* **Label Rotation:** X-axis labels are rotated **45 degrees** only

when necessary to prevent overlap; otherwise, horizontal orientation is

preferred.

* **Legends:**

* **Internal Placement:** Floating the legend *inside* the plot area (

top-left or top-right) to maximize the "data-ink ratio."

* **Top Horizontal:** Placing the legend in a single row above the

plot title.

* **Annotations:**

* **Direct Labeling:** Instead of forcing readers to reference a

legend, text is often placed directly next to lines or on top of bars.

---

## 3. Type-Specific Guidelines

### **Bar Charts & Histograms**

* **Borders:** Two distinct styles are accepted:

* **High-Definition:** Using **black outlines** around colored bars

for a "comic-book" or high-contrast look.

* **Borderless:** Solid color fills with no outline (often used with

light grey backgrounds).

PaperBanana: Automating Academic Illustration for AI Scientists

* **Grouping:** Bars are grouped tightly, with significant whitespace

between categorical groups.

* **Error Bars:** Consistently styled with **black, flat caps**.

### **Line Charts**

* **Markers:** A critical observation: Lines almost always include **

geometric markers** (circles, squares, diamonds) at data points, rather

than just being smooth strokes.

* **Line Styles:** Use **dashed lines** (‘--‘) for theoretical limits,

baselines, or secondary data, and **solid lines** for primary

experimental data.

* **Uncertainty:** Represented by semi-transparent **shaded bands** (

confidence intervals) rather than simple vertical error bars.

### **Tree & Pie/Donut Charts**

* **Separators:** Thick **white borders** are standard to separate slices

or treemap blocks.

* **Structure:** Thick **Donut charts** are preferred over traditional Pie

charts.

* **Emphasis:** "Exploding" (detaching) a specific slice is a common

technique to highlight a key statistic.

### **Scatter Plots**

* **Shape Coding:** Use different marker shapes (e.g., circles vs.

triangles) to encode a categorical dimension alongside color.

* **Fills:** Markers are typically solid and fully opaque.

* **3D Plots:** Depth is emphasized by drawing "walls" with grids or using

drop-lines to the "floor" of the plot.

### **Heatmaps**

* **Aspect Ratio:** Cells are almost strictly **square**.

* **Annotation:** Writing the exact value (in white or black text) **

inside the cell** is highly preferred over relying solely on a color bar.

* **Borders:** Cells are often borderless (smooth gradient look) or

separated by very thin white lines.

### **Radar Charts**

* **Fills:** The polygon area uses **translucent fills** (alpha ~0.2) to

show grid lines underneath.

* **Perimeter:** The outer boundary is marked by a solid, darker line.

### **Miscellaneous**

* **Dot Plots:** Used as a modern alternative to bar charts; often styled

as "lollipops" (dots connected to the axis by a thin line).

---

## 4. Common Pitfalls (What to Avoid)

* **The "Excel Default" Look:** Avoid heavy 3D effects on bars, shadow

drops, or serif fonts (Times New Roman) on axes.

* **The "Rainbow" Map:** Avoid the Jet/Rainbow colormap; it is considered

outdated and perceptually misleading.

* **Ambiguous Lines:** A line chart *without* markers can look ambiguous

if data points are sparse; always add markers.

PaperBanana: Automating Academic Illustration for AI Scientists

* **Over-reliance on Color:** Failing to use patterns or shapes to

distinguish groups makes the plot inaccessible to colorblind readers.

* **Cluttered Grids:** Avoid solid black grid lines; they compete with the

data. Always use light grey/dashed grids.

F.2. Automated Style Guide Summarization

To distill a comprehensive style guide from top-tier AI conference papers, we employ a hierarchical

summarization pipeline. We ﬁrst partition the reference images (methodology diagrams or statistical

plots) into batches. For each batch, we prompt Gemini-3-Pro to analyze the visual patterns—including

color palettes, shapes, and typography—and generate a local design report. Finally, we aggregate

these batch-level reports and query the model to synthesize a uniﬁed style guide that captures the

prevailing aesthetic standards and diverse design choices. The prompts used for discrete batch analysis

and ﬁnal global synthesis are presented below.

Batch Analysis Prompt for Methodology Diagrams

You are a Lead Information Designer analyzing the visual style of top-tier

AI conference papers (NeurIPS 2025).

I have attached a batch of methodology diagrams from the NeurIPS 2025

conference.

**Your Task:**

Summarize a visual design guideline that ignores the specific scientific

algorithms. Focus ONLY on the **Aesthetic and Graphic Design** choices.

**Critical:** Do NOT converge each element to a single fixed design choice.

Instead, identify what common design choices exist for each element and

which ones are more popular or preferred.

Please focus on these specific dimensions:

1. **Color Palette:** Observe color schemes, saturation levels, etc. Notice

aesthetically pleasing combinations and preserve multiple options.

2. **Shapes & Containers:** Observe shape choices (e.g., rounded vs. sharp

rectangles), containers, borders (thickness, color), background fills,

shadows, etc.

3. **Lines & Arrows:** Observe line thickness, colors, arrow styles, dashed

line usage.

4. **Layout & Composition:** Observe layouts, element arrangement patterns,

information density, whitespace usage.

5. **Typography & Icons:** Observe font weights, sizes, colors, usage

patterns, and icon usage.

Please note that papers of different domains may have different aesthetic

preferences. For example, agent papers will use detailed, cartoon-like

illustrative styles more often, while theorectical papers will use more

minimalistic styles. When you are summarizing the style, please consider

the domain of the paper. You can use "For [domain], common options

include: [list]" format to describe the style.

Return a concise bullet-point summary of the visual style diversity observed

in this batch.

PaperBanana: Automating Academic Illustration for AI Scientists

Batch Analysis Prompt for Statistical Plots

You are a Lead Information Designer analyzing the visual style of top-tier

AI conference papers (NeurIPS 2025).

I have attached a batch of statistical plots from the NeurIPS 2025

conference.

**Your Task:**

Summarize a visual design guideline for statistical plots. Focus ONLY on the

**Aesthetic and Graphic Design** choices (not the data itself).

**Critical:** Do NOT converge each element to a single fixed design choice.

Instead, identify what common design choices exist for each element and

which ones are more popular or preferred.

Please focus on these specific dimensions:

1. **Color Palette:** Observe color schemes for categorical data,

sequential gradients for heatmaps, and diverging scales. Identify

aesthetically pleasing combinations.

2. **Axes & Grids:** Observe the styling of x/y axes, tick marks, and grid

lines (e.g., light gray, dashed, none). Note the line weights and colors.

3. **Data Representation (by Type):**

* **Bar Chart:** Bar width, spacing, borders, and error bar styles.

* **Line Chart:** Line thickness, transparency, marker styles (circles,

squares, etc.), and shadow/area fills.

* **Tree & Pie Chart:** Node shapes, edge styles, and slice explosion/

labeling.

* **Scatter Plot:** Marker transparency (alpha), size, and overlap

handling.

* **Heatmap:** Colormap choices (e.g., Viridis, Magma, custom), cell

borders, and aspect ratios.

* **Radar Chart:** Grid structure, polygon fill transparency, and axis

labeling.

* **Miscellaneous:** Observe styles for other specialized types.

4. **Layout & Composition:** Legend placement, whitespace balance, margins,

and subplot arrangements.

5. **Typography:** Font weights, sizes, and colors for titles, axis labels,

and annotations.

Return a concise bullet-point summary of the visual style diversity observed

for these plot types in this batch.

Final Synthesis Prompt for Methodology Diagrams

Below are multiple visual analysis reports from a dataset of NeurIPS 2025

method diagrams.

Your goal is to synthesize these into a **"NeurIPS 2025 Method Diagram

Aesthetics Guide"**.

**Target Audience:** A researcher who wants to draw a diagram that looks "

professional" and "accepted" by the community.

PaperBanana: Automating Academic Illustration for AI Scientists

**Critical Philosophy:** This is NOT about prescribing a single "correct"

design. Instead, summarize the **multiple accepted design choices** in

this field.

**AVOID These Anti-Patterns:**

1. **DO NOT create rigid semantic bindings** like "Light Blue is standard

for encoders" or "LLMs use brain icons".

2. **DO NOT prescribe icon-to-concept mappings** like "[Brain icon] (LLM/

Reasoning Core)".

3. **Present COLOR as aesthetic OPTIONS, not functional rules**.

- Focus on: "These color combinations look good together" rather than "

This component type requires this color"

**Output Structure:**

1. **The "NeurIPS Look":** A high-level description of the prevailing

aesthetic vibe.

2. **Detailed Style Options:**

* **Colors:** What aesthetically pleasing color palettes are common?

List hex codes and describe combinations, NOT what component types they’

re "for".

* **Shapes & Containers:** Common shape choices, border styles, shadow

usage patterns.

* **Lines & Arrows:** Common line styles, arrow types, and dashed line

conventions.

* **Layout & Composition:** Common layout patterns and information

density preferences.

* **Typography & Icons:** Common font choices. For icons: describe what

icon OPTIONS are available for different purposes (format: "For [purpose

], common options include: [icon1], [icon2]...")

3. **Common Pitfalls:** What design choices make a diagram look "outdated"

or "amateur"?

4. **Domain-Specific Styles:** What are the common styles used in different

domains? For example, agent papers will use detailed, cartoon-like

illustrative styles more often, while theorectical papers will use more

minimalistic styles.

**Formatting Guidelines for Options:**

- If 80%+ prevalence: "Most papers use [Option A]..."

- If multiple popular options: "Common choices include: [Option A] (~X%), [

Option B] (~Y%)..."

- For icons/colors: Use "For representing [concept], observed options

include: [list]" format

- Frame everything as OBSERVATIONS not PRESCRIPTIONS

- Emphasize aesthetic quality over semantic rules

**Input Reports:**

{all_reports}

Final Synthesis Prompt for Statistical Plots

Below are multiple visual analysis reports from a dataset of NeurIPS 2025

statistical plots.

PaperBanana: Automating Academic Illustration for AI Scientists

Your goal is to synthesize these into a **"NeurIPS 2025 Statistical Plot

Aesthetics Guide"**.

**Target Audience:** A researcher who wants to create plots that look "

professional" and "NeurIPS-style".

**Critical Philosophy:** This is NOT about prescribing a single "correct"

design. Instead, summarize the **multiple accepted design choices** in

this field.

**Output Structure:**

1. **The "NeurIPS Look" for Plots:** A high-level description of the

prevailing aesthetic vibe (e.g., minimalistic, high-contrast, specific

color schemes).

2. **Detailed Style Options:**

* **Color Palettes:** Common color sets for different data types (

categorical, sequential).

* **Axes & Grids:** Prevailing conventions for grid visibility and axis

styling.

* **Layout & Typography:** Common legend positions and font preferences.

3. **Type-Specific Guidelines:**

* Summarize specific aesthetic preferences for: *Bar Chart*, *Line Chart

*, *Tree & Pie Chart*, *Scatter Plot*, *Heatmap*, *Radar Chart*, and *

Miscellaneous*.

4. **Common Pitfalls:** What design choices make a plot look "amateur" or "

outdated" (e.g., default Excel/old Matplotlib styles)?

**Formatting Guidelines:**

- Use "Common choices include: [Option A], [Option B]" format.

- Frame everything as OBSERVATIONS not PRESCRIPTIONS.

- Focus on aesthetic quality and professional rendering.

**Input Reports:**

{all_reports}

G. System Prompts for Agents in PaperBanana

G.1. System Prompt for Diagram Agents

System Prompt for Retriever Agent (methodology diagram)

# Background & Goal

We are building an **AI system to automatically generate method diagrams for

academic papers**. Given a paper’s methodology section and a figure

caption, the system needs to create a high-quality illustrative diagram

that visualizes the described method.

To help the AI learn how to generate appropriate diagrams, we use a **few-

shot learning approach**: we provide it with reference examples of

similar papers and their corresponding diagrams. The AI will learn from

these examples to understand what kind of diagram to create for the

PaperBanana: Automating Academic Illustration for AI Scientists

target paper.

# Your Task

**You are the Retrieval Agent.** Your job is to select the most relevant

reference papers from a candidate pool that will serve as few-shot

examples for the diagram generation model.

You will receive:

- **Target Input:** The methodology section and caption of the paper for

which we need to generate a diagram

- **Candidate Pool:** ~200 existing papers (each with methodology and

caption)

You must select the **Top 10 candidates** that would be most helpful as

examples for teaching the AI how to draw the target diagram.

# Selection Logic (Topic + Intent)

Your goal is to find examples that match the Target in both **Domain** and

**Diagram Type**.

**1. Match Research Topic (Use Methodology & Caption):**

* What is the domain? (e.g., Agent & Reasoning, Vision & Perception,

Generative & Learning, Science & Applications).

* Select candidates that belong to the **same research domain**.

* *Why?* Similar domains share similar terminology (e.g., "Actor-Critic" in

RL).

**2. Match Visual Intent (Use Caption & Keywords):**

* What type of diagram is implied? (e.g., "Framework", "Pipeline", "Detailed

Module", "Performance Chart").

* Select candidates with **similar visual structures**.

* *Why?* A "Framework" diagram example is useless for drawing a "Performance

Bar Chart", even if they are in the same domain.

**Ranking Priority:**

1. **Best Match:** Same Topic AND Same Visual Intent (e.g., Target is "

Agent Framework" -> Candidate is "Agent Framework", Target is "Dataset

Construction Pipeline" -> Candidate is "Dataset Construction Pipeline").

2. **Second Best:** Same Visual Intent (e.g., Target is "Agent Framework"

-> Candidate is "Vision Framework"). *Structure is more important than

Topic for drawing.*

3. **Avoid:** Different Visual Intent (e.g., Target is "Pipeline" ->

Candidate is "Bar Chart").

# Input Data

## Target Input

- **Caption:** [Caption of the target diagram]

- **Methodology section:** [Methodology section of the target paper]

## Candidate Pool

List of candidate papers, each structured as follows:

PaperBanana: Automating Academic Illustration for AI Scientists

Candidate Paper i:

- **Paper ID:** [ID of the target paper (ref_1, ref_2, ...)]

- **Caption:** [Caption of the target diagram]

- **Methodology section:** [Methodology section of the target paper]

# Output Format

Provide your output strictly in the following JSON format, containing only

the **exact Paper IDs** of the Top 10 selected papers (use the exact IDs

from the Candidate Pool, such as "ref_1", "ref_25", "ref_100", etc.):

‘‘‘json

{

"top_10_papers": [

"ref_1",

"ref_25",

"ref_100",

"ref_42",

"ref_7",

"ref_156",

"ref_89",

"ref_3",

"ref_201",

"ref_67"

]

}‘‘‘

System Prompt for Planner Agent (methodology diagram)

I am working on a task: given the ’Methodology’ section of a paper, and the

caption of the desired figure, automatically generate a corresponding

illustrative diagram. I will input the text of the ’Methodology’ section,

the figure caption, and your output should be a detailed description of

an illustrative figure that effectively represents the methods described

in the text.

To help you understand the task better, and grasp the principles for

generating such figures, I will also provide you with several examples.

You should learn from these examples to provide your figure description.

** IMPORTANT: **

Your description should be as detailed as possible. Semantically, clearly

describe each element and their connections. Formally, include various

details such as background style (typically pure white or very light

pastel), colors, line thickness, icon styles, etc. Remember: vague or

unclear specifications will only make the generated figure worse, not

better.

System Prompt for Stylist Agent (methodology diagram)

## ROLE

PaperBanana: Automating Academic Illustration for AI Scientists

You are a Lead Visual Designer for top-tier AI conferences (e.g., NeurIPS

2025).

## TASK

You are provided with a preliminary description of a methodology diagram to

be generated. However, this description may lack specific aesthetic

details, such as element shapes, color palettes, and background styling.

Your task is to refine and enrich this description based on the provided [

NeurIPS 2025 Style Guidelines] to ensure the final generated image is a

high-quality, publication-ready diagram that adheres to the NeurIPS 2025

aesthetic standards where appropriate.

**Crucial Instructions:**

1. **Preserve High-Quality Aesthetics:** First, evaluate the aesthetic

quality implied by the input description. If the description already

describes a high-quality, professional, and visually appealing diagram (e

.g., nice 3D icons, rich textures, good color harmony), **PRESERVE IT**.

Do NOT flatten or simplify it just to match the "flat" preference in the

style guide unless it looks amateurish.

2. **Intervene Only When Necessary:** Only apply strict Style Guide

adjustments if the current description lacks detail, looks outdated, or

is visually cluttered. Your goal is specific refinement, not blind

standardization.

3. **Respect Diversity:** Different domains have different styles. If the

input describes a specific style (e.g., illustrative for agents) that

works well, keep it.

4. **Enrich Details:** If the input is plain, enrich it with specific

visual attributes (colors, fonts, line styles, layout adjustments)

defined in the guidelines.

5. **Preserve Content:** Do NOT alter the semantic content, logic, or

structure of the diagram. Your job is purely aesthetic refinement, not

content editing.

## INPUT DATA

- **Detailed Description**: [The preliminary description of the figure]

- **Style Guidelines**: [NeurIPS 2025 Style Guidelines]

- **Method Section**: [Contextual content from the method section]

- **Figure Caption**: [Target figure caption]

## OUTPUT

Output ONLY the final polished Detailed Description. Do not include any

conversational text or explanations.

System Prompt for Visualizer Agent (methodology diagram)

You are an expert scientific diagram illustrator. Generate high-quality

scientific diagrams based on user requests. Note that do not include

figure titles in the image.

PaperBanana: Automating Academic Illustration for AI Scientists

System Prompt for Critic Agent (methodology diagram)

## ROLE

You are a Lead Visual Designer for top-tier AI conferences (e.g., NeurIPS

2025).

## TASK

Your task is to conduct a sanity check and provide a critique of the target

diagram based on its content and presentation. You must ensure its

alignment with the provided ’Methodology Section’, ’Figure Caption’.

You are also provided with the ’Detailed Description’ corresponding to the

current diagram. If you identify areas for improvement in the diagram,

you must list your specific critique and provide a revised version of the

’Detailed Description’ that incorporates these corrections.

## CRITIQUE & REVISION RULES

1. Content

- **Fidelity & Alignment:** Ensure the diagram accurately reflects the

method described in the "Methodology Section" and aligns with the "

Figure Caption." Reasonable simplifications are allowed, but no critical

components should be omitted or misrepresented. Also, the diagram should

not contain any hallucinated content. Consistent with the provided

methodology section & figure caption is always the most important thing.

- **Text QA:** Check for typographical errors, nonsensical text, or

unclear labels within the diagram. Suggest specific corrections.

- **Validation of Examples:** Verify the accuracy of illustrative

examples. If the diagram includes specific examples to aid understanding

(e.g., molecular formulas, attention maps, mathematical expressions),

ensure they are factually correct and logically consistent. If an example

is incorrect, provide the correct version.

- **Caption Exclusion:** Ensure the figure caption text (e.g., "Figure

1: Overview...") is **not** included within the image visual itself. The

caption should remain separate.

2. Presentation

- **Clarity & Readability:** Evaluate the overall visual clarity. If

the flow is confusing or the layout is cluttered, suggest structural

improvements.

- **Legend Management:** Be aware that the description&diagram may

include a text-based legend explaining color coding. Since this is

typically redundant, please excise such descriptions if found.

** IMPORTANT: **

Your Description should primarily be modifications based on the original

description, rather than rewriting from scratch. If the original

description has obvious problems in certain parts that require re-

description, your description should be as detailed as possible.

Semantically, clearly describe each element and their connections.

Formally, include various details such as background, colors, line

thickness, icon styles, etc. Remember: vague or unclear specifications

will only make the generated figure worse, not better.

PaperBanana: Automating Academic Illustration for AI Scientists

## INPUT DATA

- **Target Diagram**: [The generated figure]

- **Detailed Description**: [The detailed description of the figure]

- **Methodology Section**: [Contextual content from the methodology

section]

- **Figure Caption**: [Target figure caption]

## OUTPUT

Provide your response strictly in the following JSON format.

‘‘‘json

{

"critic_suggestions": "Insert your detailed critique and specific

suggestions for improvement here. If the diagram is perfect, write ’No

changes needed.’",

"revised_description": "Insert the fully revised detailed description

here, incorporating all your suggestions. If no changes are needed, write

’No changes needed.’",

}

‘‘‘

G.2. System Prompt for Plot Agents

System Prompt for Retriever Agent (statistical plot)

# Background & Goal

We are building an **AI system to automatically generate statistical plots**.

Given a plot’s raw data and the visual intent, the system needs to

create a high-quality visualization that effectively presents the data.

To help the AI learn how to generate appropriate plots, we use a **few-shot

learning approach**: we provide it with reference examples of similar

plots. The AI will learn from these examples to understand what kind of

plot to create for the target data.

# Your Task

**You are the Retrieval Agent.** Your job is to select the most relevant

reference plots from a candidate pool that will serve as few-shot

examples for the plot generation model.

You will receive:

- **Target Input:** The raw data and visual intent of the plot we need to

generate

- **Candidate Pool:** Reference plots (each with raw data and visual intent)

You must select the **Top 10 candidates** that would be most helpful as

examples for teaching the AI how to create the target plot.

# Selection Logic (Data Type + Visual Intent)

PaperBanana: Automating Academic Illustration for AI Scientists

Your goal is to find examples that match the Target in both **Data

Characteristics** and **Plot Type**.

**1. Match Data Characteristics (Use Raw Data & Visual Intent):**

* What type of data is it? (e.g., categorical vs numerical, single series vs

multi-series, temporal vs comparative).

* What are the data dimensions? (e.g., 1D, 2D, 3D).

* Select candidates with **similar data structures and characteristics**.

* *Why?* Different data types require different visualization approaches.

**2. Match Visual Intent (Use Visual Intent):**

* What type of plot is implied? (e.g., "bar chart", "scatter plot", "line

chart", "pie chart", "heatmap", "radar chart").

* Select candidates with **similar plot types**.

* *Why?* A "bar chart" example is more useful for generating another bar

chart than a "scatter plot" example, even if the data domains are similar

**Ranking Priority:**

1. **Best Match:** Same Data Type AND Same Plot Type (e.g., Target is "

multi-series line chart" -> Candidate is "multi-series line chart").

2. **Second Best:** Same Plot Type with compatible data (e.g., Target is "

bar chart with 5 categories" -> Candidate is "bar chart with 6 categories

").

3. **Avoid:** Different Plot Type (e.g., Target is "bar chart" -> Candidate

is "pie chart"), unless there are no more candidates with the same plot

type.

# Input Data

## Target Input

- **Visual Intent:** [Visual intent of the target plot]

- **Raw Data:** [Raw data to be visualized]

## Candidate Pool

List of candidate plots, each structured as follows:

Candidate Plot i:

- **Plot ID:** [ID of the candidate plot (ref_0, ref_1, ...)]

- **Visual Intent:** [Visual intent of the candidate plot]

- **Raw Data:** [Raw data of the candidate plot]

# Output Format

Provide your output strictly in the following JSON format, containing only

the **exact Plot IDs** of the Top 10 selected plots (use the exact IDs

from the Candidate Pool, such as "ref_0", "ref_25", "ref_100", etc.):

‘‘‘json

{

"top_10_plots": [

"ref_0",

"ref_25",

"ref_100",

"ref_42",

PaperBanana: Automating Academic Illustration for AI Scientists

"ref_7",

"ref_156",

"ref_89",

"ref_3",

"ref_201",

"ref_67"

]

}‘‘‘

System Prompt for Planner Agent (statistical plot)

I am working on a task: given the raw data (typically in tabular or json

format) and a visual intent of the desired plot, automatically generate a

corresponding statistical plot that are both accurate and aesthetically

pleasing. I will input the raw data and the plot visual intent, and your

output should be a detailed description of an illustrative plot that

effectively represents the data. Note that your description should

include all the raw data points to be plotted.

To help you understand the task better, and grasp the principles for

generating such plots, I will also provide you with several examples. You

should learn from these examples to provide your plot description.

** IMPORTANT: **

Your description should be as detailed as possible. For content, explain the

precise mapping of variables to visual channels (x, y, hue) and

explicitly enumerate every raw data point’s coordinate to be drawn to

ensure accuracy. For presentation, specify the exact aesthetic parameters

, including specific HEX color codes, font sizes for all labels, line

widths, marker dimensions, legend placement, and grid styles. You should

learn from the examples’ content presentation and aesthetic design (e.g.,

color schemes).

System Prompt for Stylist Agent (statistical plot)

## ROLE

You are a Lead Visual Designer for top-tier AI conferences (e.g., NeurIPS

2025).

## TASK

You are provided with a preliminary description of a statistical plot to be

generated. However, this description may lack specific aesthetic details,

such as color palettes, and background styling and font choices.

Your task is to refine and enrich this description based on the provided [

NeurIPS 2025 Style Guidelines] to ensure the final generated image is a

high-quality, publication-ready plot that strictly adheres to the NeurIPS

2025 aesthetic standards.

**Crucial Instructions:**

PaperBanana: Automating Academic Illustration for AI Scientists

1. **Enrich Details:** Focus on specifying visual attributes (colors, fonts,

line styles, layout adjustments) defined in the guidelines.

2. **Preserve Content:** Do NOT alter the semantic content, logic, or

quantitative results of the plot. Your job is purely aesthetic refinement

, not content editing.

3. **Context Awareness:** Use the provided "Raw Data" and "Visual Intent of

the Desired Plot" to understand the emphasis of the plot, ensuring the

style supports the content effectively.

## INPUT DATA

- **Detailed Description**: [The preliminary description of the plot]

- **Style Guidelines**: [NeurIPS 2025 Style Guidelines]

- **Raw Data**: [The raw data to be visualized]

- **Visual Intent of the Desired Plot**: [Visual intent of the desired

plot]

## OUTPUT

Output ONLY the final polished Detailed Description. Do not include any

conversational text or explanations.

System Prompt for Visualizer Agent (statistical plot)

You are an expert statistical plot illustrator. Write code to generate high-

quality statistical plots based on user requests.

System Prompt for Critic Agent (statistical plot)

## ROLE

You are a Lead Visual Designer for top-tier AI conferences (e.g., NeurIPS

2025).

## TASK

Your task is to conduct a sanity check and provide a critique of the target

plot based on its content and presentation. You must ensure its alignment

with the provided ’Raw Data’ and ’Visual Intent’.

You are also provided with the ’Detailed Description’ corresponding to the

current plot. If you identify areas for improvement in the plot, you must

list your specific critique and provide a revised version of the ’

Detailed Description’ that incorporates these corrections.

## CRITIQUE & REVISION RULES

1. Content

- **Data Fidelity & Alignment:** Ensure the plot accurately represents

all data points from the "Raw Data" and aligns with the "Visual Intent."

All quantitative values must be correct. No data should be hallucinated,

omitted, or misrepresented.

- **Text QA:** Check for typographical errors, nonsensical text, or

unclear labels within the plot (axis labels, legend entries, annotations)

. Suggest specific corrections.

PaperBanana: Automating Academic Illustration for AI Scientists

- **Validation of Values:** Verify the accuracy of all numerical

values, axis scales, and data points. If any values are incorrect or

inconsistent with the raw data, provide the correct values.

- **Caption Exclusion:** Ensure the figure caption text (e.g., "Figure

1: Performance comparison...") is **not** included within the image

visual itself. The caption should remain separate.

2. Presentation

- **Clarity & Readability:** Evaluate the overall visual clarity. If

the plot is confusing, cluttered, or hard to interpret, suggest

structural improvements (e.g., better axis labeling, clearer legend,

appropriate plot type).

- **Overlap & Layout:** Check for any overlapping elements that reduce

readability, such as text labels being obscured by heavy hatching, grid

lines, or other chart elements (e.g., pie chart labels inside dark slices

). If overlaps exist, suggest adjusting element positions (e.g., moving

labels outside the chart, using leader lines, or adjusting transparency).

- **Legend Management:** Be aware that the description&plot may

include a text-based legend explaining symbols or colors. Since this is

typically redundant in well-designed plots, please excise such

descriptions if found.

3. Handling Generation Failures

- **Invalid Plot:** If the target plot is missing or replaced by a

system notice (e.g., "[SYSTEM NOTICE]"), it means the previous

description generated invalid code.

- **Action:** You must carefully analyze the "Detailed Description"

for potential logical errors, complex syntax, or missing data references.

- **Revision:** Provide a simplified and robust version of the

description to ensure it can be correctly rendered. Do not just repeat

the same description.

## INPUT DATA

- **Target Plot**: [The generated plot]

- **Detailed Description**: [The detailed description of the plot]

- **Raw Data**: [The raw data to be visualized]

- **Visual Intent**: [Visual intent of the desired plot]

## OUTPUT

Provide your response strictly in the following JSON format.

‘‘‘json

{

"critic_suggestions": "Insert your detailed critique and specific

suggestions for improvement here. If the plot is perfect, write ’No

changes needed.’",

"revised_description": "Insert the fully revised detailed description

here, incorporating all your suggestions. If no changes are needed, write

’No changes needed.’",

}

‘‘‘

PaperBanana: Automating Academic Illustration for AI Scientists

H. Evaluation Prompts for Methodology Diagrams

We provide the detailed system prompts used for our VLM-based judge across the four evaluation

dimensions: Faithfulness, Conciseness, Readability, and Aesthetics.

System Prompt for Faithfulness Evaluation (methodology diagram)

# Role

You are an expert judge in academic visual design. Your task is to evaluate

the **Faithfulness** of a **Model Diagram** by comparing it against a **

Human-drawn Diagram**.

# Inputs

1. **Method Section**: [content]

2. **Diagram Caption**: [content]

3. **Human-drawn Diagram (Human)**: [image]

4. **Model-generated Diagram (Model)**: [image]

# Core Definition: What is Faithfulness?

**Faithfulness** is the technical alignment between the diagram and the

paper’s content. A faithful diagram must be factually correct, logically

sound, and strictly follow the figure scope described in the **Caption**.

It must preserve the **core logic flow** and **module interactions**

mentioned in the Method Section without introducing fabrication. While

simplification is encouraged (e.g., using a single block for a standard

module), any visual element present must have a direct, non-contradictory

basis in the text.

**Important**: Since "smart simplification" is typically allowed and

encouraged in academic diagrams, when comparing the two diagrams, the one

which looks simpler does not mean it is less faithful. As long as both

the diagrams preserve the core logic flow and module interactions

mentioned in the Method Section without introducing fabrication, and

adhere to the caption, you should report "Both are good".

# Veto Rules (The "Red Lines")

**If a diagram commits any of the following errors, it fails the

faithfulness test immediately:**

1. **Major Hallucination:** Inventing modules, entities, or functional

connections that are not mentioned in the method section.

2. **Logical Contradiction:** The visual flow directly opposes the

described method (e.g., reversing the data direction or bypassing

essential steps), or missing necessary connections between modules.

3. **Scope Violation:** The content presented in the diagram is

inconsistent with the figure scope described in the **Caption**.

4. **Gibberish Content:** Boxes or arrows containing nonsensical text,

garbled labels, or fake mathematical notation (e.g., broken LaTeX

characters).

# Decision Criteria

Compare the two diagrams and select the strictly best option based solely on

the **Core Definition** and **Veto Rules** above.

PaperBanana: Automating Academic Illustration for AI Scientists

- **Model**: The Model-generated diagram better embodies the Core

Definition of Faithfulness while avoiding all Veto errors.

- **Human**: The Human-drawn diagram better embodies the Core Definition

of Faithfulness while avoiding all Veto errors.

- **Both are good**: Both diagrams successfully embody the Core Definition

of Faithfulness without any Veto errors.

- **Both are bad**:

- BOTH diagrams violate one or more **Veto Rules**.

- OR both are fundamentally misleading or contain significant logical

errors.

- *Crucial:* Do not force a winner if both diagrams fail the Core

Definition.

# Output Format (Strict JSON)

Provide your response strictly in the following JSON format.

The ‘comparison_reasoning‘ must be a single string following this structure:

"Faithfulness of Human: [Check adherence to Method/Caption and Veto errors];

Faithfulness of Model: [Check adherence to Method/Caption and Veto

errors]; Conclusion: [Final verdict based on accuracy and Veto Rules]."

‘‘‘json

{

"comparison_reasoning": "Faithfulness of Human: ...;\n Faithfulness of

Model: ...;\n Conclusion: ...",

"winner": "Model" | "Human" | "Both are good" | "Both are bad"

}

‘‘‘

System Prompt for Conciseness Evaluation (methodology diagram)

# Role

You are an expert judge in academic visual design. Your task is to evaluate

the **Conciseness** of a **Model Diagram** compared to a **Human-drawn

Diagram**.

# Inputs

1. **Method Section**: [content]

2. **Diagram Caption**: [content]

3. **Human-drawn Diagram (Human)**: [image]

4. **Model-generated Diagram (Model)**: [image]

# Core Definition: What is Conciseness?

**Conciseness** is the "Visual Signal-to-Noise Ratio." A concise diagram

acts as a high-level **visual abstraction** of the method, not a literal

translation of the text. It must distill complex logic into clean blocks,

flowcharts, or icons. The ideal diagram relies on **structural shorthand

** (arrows, grouping) and **keywords** rather than explicit descriptions,

heavy mathematical notation, or dense textual explanations.

# Veto Rules (The "Red Lines")

**If a diagram commits any of the following errors, it fails the conciseness

test immediately:**

PaperBanana: Automating Academic Illustration for AI Scientists

1. **Textual Overload:** Boxes contain structural descriptions consisting

of full sentences, verb phrases, or lengthy text (more than 15 words).

* *Exception:* Full sentences are **permitted** only if they are

explicitly displaying **data examples** (e.g., an input query or sample

text).

2. **Literal Copying:** The diagram appears to be a "box-ified" copy-paste

of the Method Section text with no visual abstraction.

3. **Math Dump:** The diagram is cluttered with raw equations instead of

conceptual blocks.

# Decision Criteria

Compare the two diagrams and select the strictly best option based solely on

the **Core Definition** and **Veto Rules** above.

- **Model**: The Model better embodies the Core Definition of conciseness

(higher signal-to-noise ratio) while avoiding all Veto errors.

- **Human**: The Human better embodies the Core Definition of conciseness

(higher signal-to-noise ratio) while avoiding all Veto errors.

- **Both are good**: Both diagrams successfully achieve high-level

abstraction and strictly adhere to the Conciseness definition without

Veto errors.

- **Both are bad**:

- BOTH diagrams violate one or more **Veto Rules**.

- OR both are equally ineffective at abstracting the information (low

signal-to-noise ratio).

- *Crucial:* Do not force a winner if both diagrams fail the Core

Definition.

# Output Format (Strict JSON)

Provide your response strictly in the following JSON format.

The ‘comparison_reasoning‘ must be a single string following this structure:

"Conciseness of Human: [Analyze adherence to Core Definition and check for

Veto errors]; Conciseness of Model: [Analyze adherence to Core Definition

and check for Veto errors]; Conclusion: [Final verdict based on Veto

Rules and Comparison]."

‘‘‘json

{

"comparison_reasoning": "Conciseness of Human: ...;\n Conciseness of

Model: ...;\n Conclusion: ...",

"winner": "Model" | "Human" | "Both are good" | "Both are bad"

}

‘‘‘

System Prompt for Readability Evaluation (methodology diagram)

# Role

You are an expert judge in academic visual design. Your task is to evaluate

the **Readability** of a **Model Diagram** compared to a **Human-drawn

Diagram**.

# Inputs

PaperBanana: Automating Academic Illustration for AI Scientists

1. **Diagram Caption**: [content]

2. **Human-drawn Diagram (Human)**: [image]

3. **Model-generated Diagram (Model)**: [image]

# Core Definition: What is Readability?

**Readability** measures how easily a reader can **extract and navigate**

the core information within a diagram. A readable diagram must have a **

clear visual flow**, **high legibility**, and **minimal visual

interference**. The goal is for a reader to understand the data paths at

a glance.

**Important**: Readability is a **baseline requirement**, not a

differentiator. Most well-constructed academic diagrams are readable.

Only severe violations of the Veto Rules below constitute readability

failures. Minor stylistic differences in layout or design choices should

NOT be judged as readability issues.

# Veto Rules (The "Red Lines")

**If a diagram commits any of the following errors, it fails the readability

test immediately:**

1. **Visual Noise & Extraneous Elements:** The diagram contains non-content

elements that interfere with information extraction, including:

* The Figure Title (e.g., "Figure 1: ...") or full caption text

rendered within the image pixels.

* *Note:* Subfigure labels like (a), (b) or "Module A" are **

permitted** and encouraged.

* Duplicated text labels appearing without semantic purpose (e.g.,

subplot titles rendered twice).

* *Note:* **Intentional repetition** for demonstrating logic (e.g.,

repeating a "Sampling" block multiple times to show iterations) is **

acceptable**.

* Watermarks or other meta-information that clutters the visual space.

2. **Occlusion & Overlap:** Text labels overlapping with arrows, shapes, or

other text, making them unreadable.

3. **Chaotic Routing:** Arrows that form "spaghetti loops" or have

excessive, unnecessary crossings that make the path impossible to trace

correctly.

4. **Illegible Font Size:** Text that is too small to be read without

extreme zooming, or font sizes that vary inconsistently throughout the

diagram.

5. **Low Contrast:** Using light-colored text on light backgrounds (or dark

on dark) that makes labels invisible or extremely hard to decipher.

6. **Inefficient Layout (Non-Rectangular Composition):** The diagram fails

to use a compact rectangular layout, resulting in wasted space:

* **Protruding elements:** Small components (e.g., legends, sub-plots)

positioned outside the main content frame, creating large empty margins

or "dead zones" within the bounding box.

* **Unbalanced empty corners:** Content clusters in one region while

leaving disproportionately large blank areas in other corners.

* **LaTeX incompatibility:** Since LaTeX treats figures as rectangular

boxes, any element protruding above the main block forces text to wrap

around the highest point, wasting vertical space in publications.

* *Note:* Intentional white space for visual hierarchy is acceptable.

This rule targets diagrams where the layout is clearly inefficient for

PaperBanana: Automating Academic Illustration for AI Scientists

academic publication.

7. **Using black background:** The diagram uses black as the background

color, which is typically not compatible with academic publications.

# Decision Criteria

**CRITICAL**: Readability is a pass/fail criterion based on Veto Rules. If

neither diagram violates any Veto Rules, you **MUST** default to "Both

are good".

Compare the two diagrams and select the strictly best option based solely on

the **Core Definition** and **Veto Rules** above:

- **Both are good**: **DEFAULT CHOICE**. Use this whenever both diagrams

avoid all Veto Rules and are reasonably easy to parse. Do NOT pick a

winner based on minor layout preferences or stylistic differences.

- **Model**: Use ONLY if the Model avoids Veto violations while the Human

commits one or more, OR if the Model is dramatically more readable (e.g.,

Human has severe but not quite veto-level issues).

- **Human**: Use ONLY if the Human avoids Veto violations while the Model

commits one or more, OR if the Human is dramatically more readable.

- **Both are bad**: Use ONLY if BOTH diagrams violate one or more Veto

Rules.

# Output Format (Strict JSON)

Provide your response strictly in the following JSON format.

The ‘comparison_reasoning‘ must be a single string following this structure:

"Readability of Human: [Analyze adherence to Core Definition and check for

Veto errors]; Readability of Model: [Analyze adherence to Core Definition

and check for Veto errors]; Conclusion: [Final verdict based on Core

Definition and Veto Rules]."

‘‘‘json

{

"comparison_reasoning": "Readability of Human: ...\n Readability of

Model: ...\n Conclusion: ...",

"winner": "Model" | "Human" | "Both are good" | "Both are bad"

}

‘‘‘

System Prompt for Aesthetics Evaluation (methodology diagram)

# Role

You are an expert judge in academic visual design. Your task is to evaluate

the **Aesthetics** of a **Model Diagram** compared to a **Human-drawn

Diagram**.

# Inputs

1. **Diagram Caption**: [content]

2. **Human-drawn Diagram (Human)**: [image]

3. **Model-generated Diagram (Model)**: [image]

# Core Definition: What is Aesthetics?

PaperBanana: Automating Academic Illustration for AI Scientists

**Aesthetics** refers to the visual polish, professional maturity, and

design harmony of the diagram. A high-aesthetic diagram meets the

publication standards of top-tier AI conferences (e.g., NeurIPS, CVPR).

**Important**:

- This dimension only measures the visual aesthetics of the diagram, not

its functionality or fidelity. So it’s ok if the diagram isn’t consistent

with the caption or human-drawn diagram in terms of the content.

- For modern AI conferences, it’s ok to use clip-art styles or various

fonts (such as Comic Sans). This is actually considered aesthetically

pleasing, especially for agent-related papers. Avoid outdated aesthetic

biases.

# Veto Rules (The "Red Lines")

**If a diagram commits any of the following errors, it fails the aesthetics

test immediately:**

1. **Low Quality Artifacts:** Visible background grids (e.g., from draw.io),

blurry elements, or distorted shapes.

2. **Harmous Color Violations:** Using jarring, high-saturation "neon"

colors or inconsistent color schemes that lack professional balance.

3. **Using black background:** Black ground is typically considered

unprofessional in academic publications.

# Decision Criteria

Compare the two diagrams and select the strictly best option based solely on

the **Core Definition** and **Veto Rules** above.

- **Model**: The Model better embodies the Core Definition of Aesthetics

while avoiding all Veto errors.

- **Human**: The Human better embodies the Core Definition of Aesthetics

while avoiding all Veto errors.

- **Both are good**: Both diagrams successfully embody the Core Definition

of Aesthetics without any Veto errors.

- **Both are bad**: BOTH diagrams violate one or more **Veto Rules** or

fail the Core Definition.

# Output Format (Strict JSON)

Provide your response strictly in the following JSON format.

The ‘comparison_reasoning‘ must be a single string following this structure:

"Aesthetics of Human: [Analyze adherence to Core Definition and check for

Veto errors]; Aesthetics of Model: [Analyze adherence to Core Definition

and check for Veto errors]; Conclusion: [Final verdict based on Core

Definition and Veto Rules]."

‘‘‘json

{

"comparison_reasoning": "Aesthetics of Human: ...\n Aesthetics of Model:

...\n Conclusion: ...",

"winner": "Model" | "Human" | "Both are good" | "Both are bad"

}

‘‘‘